Blind Separation of Speech Mixtures

57
Blind Separation of Speech Mixtures Vaninirappuputhenpurayil Gopalan REJU School of Electrical and Electronic Engineering Nanyang Technological University 02:45 AM 1

description

Blind Separation of Speech Mixtures. Vaninirappuputhenpurayil Gopalan REJU School of Electrical and Electronic Engineering Nanyang Technological University. Introduction. Blind Source Separation. Convolutive. Mixing process:. s 1. s 2. Unmixing process:. Introduction. - PowerPoint PPT Presentation

Transcript of Blind Separation of Speech Mixtures

Page 1: Blind Separation of Speech Mixtures

Blind Separation of Speech MixturesBlind Separation of Speech Mixtures

Vaninirappuputhenpurayil Gopalan REJU

School of Electrical and Electronic EngineeringNanyang Technological University

Vaninirappuputhenpurayil Gopalan REJU

School of Electrical and Electronic EngineeringNanyang Technological University

0510 AM 1

Introduction Introduction

Blind Source Separation Blind Source Separation

0510 AM

0

2120

1111 )()()()()(ll

ltslhltslhtx

0

2220

1212 )()()()()(ll

ltslhltslhtx

L

l

L

l

ltxlwltxlwty0

2120

1111 )()()()()(

L

l

L

l

ltxlwltxlwty0

2220

1212 )()()()()(

bull Mixing process

bull Unmixing process

ConvolutiveConvolutive

2

s1

s2

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

0510 AM 3

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

tx

tx

hh

hh

)()( tt SX H

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

hh

hh

tx

tx

)()( tt HSX

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

tfS

tfS

fHfH

fHfH

tfX

tfX

)()()( tfftf SX H

bull In frequency domain

Difficu

lt to

separa

te

Difficu

lt to

separa

te

Easy to

separa

te

Easy to

separa

te

0510 AM 4

Introduction Introduction

No of sources lt No of sensorNo of sources lt No of sensor1s

2s

3s

4s

1x2x3x

1s

2s1x2x3x

3s

1s

2s 1x2x3x

4s

No of sources = No of sensorNo of sources = No of sensor

No of sources gt No of sensorNo of sources gt No of sensor

Overdetermined mixingOverdetermined mixing

Determined mixingDetermined mixing

Underdetermined mixingUnderdetermined mixing

Difficult to se

parate

Difficult to se

parate

Easy to se

parate

Easy to se

parate

0510 AM 5

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 2: Blind Separation of Speech Mixtures

Introduction Introduction

Blind Source Separation Blind Source Separation

0510 AM

0

2120

1111 )()()()()(ll

ltslhltslhtx

0

2220

1212 )()()()()(ll

ltslhltslhtx

L

l

L

l

ltxlwltxlwty0

2120

1111 )()()()()(

L

l

L

l

ltxlwltxlwty0

2220

1212 )()()()()(

bull Mixing process

bull Unmixing process

ConvolutiveConvolutive

2

s1

s2

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

0510 AM 3

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

tx

tx

hh

hh

)()( tt SX H

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

hh

hh

tx

tx

)()( tt HSX

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

tfS

tfS

fHfH

fHfH

tfX

tfX

)()()( tfftf SX H

bull In frequency domain

Difficu

lt to

separa

te

Difficu

lt to

separa

te

Easy to

separa

te

Easy to

separa

te

0510 AM 4

Introduction Introduction

No of sources lt No of sensorNo of sources lt No of sensor1s

2s

3s

4s

1x2x3x

1s

2s1x2x3x

3s

1s

2s 1x2x3x

4s

No of sources = No of sensorNo of sources = No of sensor

No of sources gt No of sensorNo of sources gt No of sensor

Overdetermined mixingOverdetermined mixing

Determined mixingDetermined mixing

Underdetermined mixingUnderdetermined mixing

Difficult to se

parate

Difficult to se

parate

Easy to se

parate

Easy to se

parate

0510 AM 5

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 3: Blind Separation of Speech Mixtures

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

0510 AM 3

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

tx

tx

hh

hh

)()( tt SX H

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

hh

hh

tx

tx

)()( tt HSX

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

tfS

tfS

fHfH

fHfH

tfX

tfX

)()()( tfftf SX H

bull In frequency domain

Difficu

lt to

separa

te

Difficu

lt to

separa

te

Easy to

separa

te

Easy to

separa

te

0510 AM 4

Introduction Introduction

No of sources lt No of sensorNo of sources lt No of sensor1s

2s

3s

4s

1x2x3x

1s

2s1x2x3x

3s

1s

2s 1x2x3x

4s

No of sources = No of sensorNo of sources = No of sensor

No of sources gt No of sensorNo of sources gt No of sensor

Overdetermined mixingOverdetermined mixing

Determined mixingDetermined mixing

Underdetermined mixingUnderdetermined mixing

Difficult to se

parate

Difficult to se

parate

Easy to se

parate

Easy to se

parate

0510 AM 5

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 4: Blind Separation of Speech Mixtures

Introduction Introduction

Convolutive Blind Source Separation Convolutive Blind Source Separation Instantaneous Blind Source Separation Instantaneous Blind Source Separation

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

tx

tx

hh

hh

)()( tt SX H

)(

)(

)(

)(

2

1

2221

1211

2

1

ts

ts

hh

hh

tx

tx

)()( tt HSX

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

tfS

tfS

fHfH

fHfH

tfX

tfX

)()()( tfftf SX H

bull In frequency domain

Difficu

lt to

separa

te

Difficu

lt to

separa

te

Easy to

separa

te

Easy to

separa

te

0510 AM 4

Introduction Introduction

No of sources lt No of sensorNo of sources lt No of sensor1s

2s

3s

4s

1x2x3x

1s

2s1x2x3x

3s

1s

2s 1x2x3x

4s

No of sources = No of sensorNo of sources = No of sensor

No of sources gt No of sensorNo of sources gt No of sensor

Overdetermined mixingOverdetermined mixing

Determined mixingDetermined mixing

Underdetermined mixingUnderdetermined mixing

Difficult to se

parate

Difficult to se

parate

Easy to se

parate

Easy to se

parate

0510 AM 5

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 5: Blind Separation of Speech Mixtures

Introduction Introduction

No of sources lt No of sensorNo of sources lt No of sensor1s

2s

3s

4s

1x2x3x

1s

2s1x2x3x

3s

1s

2s 1x2x3x

4s

No of sources = No of sensorNo of sources = No of sensor

No of sources gt No of sensorNo of sources gt No of sensor

Overdetermined mixingOverdetermined mixing

Determined mixingDetermined mixing

Underdetermined mixingUnderdetermined mixing

Difficult to se

parate

Difficult to se

parate

Easy to se

parate

Easy to se

parate

0510 AM 5

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 6: Blind Separation of Speech Mixtures

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Types of mixing

Instantaneous mixing Convolutive mixing

0510 AM 6

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 7: Blind Separation of Speech Mixtures

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixing

Step 1 Selection of cost function

Step 2 Minimization or maximization of the cost function

0510 AM

WHS1

S2 X2

Y1

Y2

Separated

X1

7

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 8: Blind Separation of Speech Mixtures

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsInstantaneous mixingInstantaneous mixing

Selection of cost function

Statistical independence

Information theoretic

Non-Gaussianity

Kurtosis

Negentropy

Nonlinear cross moments

Temporal structure of speech

Non-stationarity of speech0510 AM

i

ii ypp y is idea Basic

Central limit theoremMixture of two or more sources will be more Gaussian than their individual components

224 E3E)( yyykurt

yyy HHJ gauss

yyyy dppH logEntropy

0 ji ygyfE

used becan statsticsorder Secondlags timedifferent for usly simultaneon correlatiooutput theedioganaliz eg

eddiagonalizusly simultaneo are matricesn correlatio the

and blocks into divided are Signals

Non Gaussianity measures

Signals from two different sources are independent

8

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 9: Blind Separation of Speech Mixtures

Approaches for BSS of Speech SignalsApproaches for BSS of Speech Signals

Instantaneous mixingMinimization or maximization of the cost function

simple gradient method

Natural gradient method

Newtonrsquos method

eg Informax ICA algorithm

eg FastICA

0510 AM 9

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 10: Blind Separation of Speech Mixtures

Approaches for BSS of Speech SignalsApproaches for BSS of Speech SignalsConvolutive Mixing

Time Domain Frequency Domain

QqltxlwyP

p

L

lpqpq 1)()(

1

1

0

)()()( tfftf SHX

)()()( tfftf XWY

AdvantageNo permutation problem

DisadvantageSlow convergence High computational cost for long filter taps

AdvantageLow computational costFast convergence

DisadvantagePermutation Problem

WHS1

S2

X1

X2

Y1 Y2

Y2 Y1

0510 AM 10

or

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 11: Blind Separation of Speech Mixtures

Permutation Problem in Frequency Domain BSS

Permutation Problem in Frequency Domain BSS

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Still signals are mixed

K po

int

IFFT

Corresponding to different sources Due to permutation problem

One frequency binInstantaneous ICA algorithm

Solv

ing

perm

utati

on

Prob

lem y1

y2y3

Separated signals

Corresponding to y3

0510 AM 11

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 12: Blind Separation of Speech Mixtures

MotivationMotivation

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

12

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 13: Blind Separation of Speech Mixtures

My Contribution - IMy Contribution - I

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

13

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 14: Blind Separation of Speech Mixtures

Algorithm for Solving the Permutation Problem

Algorithm for Solving the Permutation Problem

f1

f2

fk

x1x2

x3

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Permutation problem

One frequency binInstantaneous ICA algorithm

Permutation problem solved

0510 AM 14

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 15: Blind Separation of Speech Mixtures

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Direction Of Arrival (DOA) method

2

1

)sin(2 1

p

dcfjkqpkq

pkefWfU Position of the pth sensor

Velocity of sound

)(

)(

)()(

)()(

)(

)(

2

1

2221

1211

2

1

k

k

kk

kk

k

k

fX

fX

fWfW

fWfW

fY

fY

0510 AM

Direction of y1 = -30o

Direction of y2 = 20o

15

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 16: Blind Separation of Speech Mixtures

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Reasons for failure at lower freq

Lower spacing causes error in phase difference measurement

The relation is approximated for plane wave front under anechoic condition

Disadvantages

Fails at lower frequenciesFails when sources are nearRoom reverberationSensor positions must be known

Direction Of Arrival (DOA) method

0510 AM 16

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 17: Blind Separation of Speech Mixtures

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

f1

f2

fk

BSS

BSS

BSSMixed signals

K po

int

FFT

y1y2y3

Separated signals

K po

int

IFFT

Solv

ing

perm

utati

on

Prob

lem

Low correlation

High correlationLow

correlation

x1x2

x3

Adjacent bands correlation method

0510 AM 17

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 18: Blind Separation of Speech Mixtures

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21r11 r12

r21 r22

s1

s2

Correlation matrix

21122211 rrrr No change

21122211 rrrr Change permutation

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM

21122211 and and rrrr

21122211 and and rrrr

With confidence Without confidence

2090

8010

8020

1090

2090

1040

1040

2090

Example Example

18

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 19: Blind Separation of Speech Mixtures

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

K-1

K

K+1 K+2

K+3

helliphelliphelliphellip

r12

r21

r11

r22 r22 r22 r22

r11 r11 r11

r12

r21

r12

r21

r12

r21

r11 r12

r21 r22

s1

s2

Correlation matrix

Disadvantage The method is not robust

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Adjacent bands correlation method

0510 AM 19

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 20: Blind Separation of Speech Mixtures

0510 AM

Existing Method forSolving the Permutation Problem

Existing Method forSolving the Permutation Problem

Combination of DOA and Correlation methods method

DOA + Harmonic Correlation + Adjacent bands correlation

Advantage Increased robustness

20

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 21: Blind Separation of Speech Mixtures

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

Proposed algorithm Partial separation method(Parallel configuration)

Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112

2y

1y

1x

2x

1s

2s

0510 AM 21

Time domain stage

Frequency domain stage

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 22: Blind Separation of Speech Mixtures

Partial separation method(Parallel configuration)

Partial separation method(Parallel configuration)

0510 AM 22

Time domain stage

Frequency domain stage

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 23: Blind Separation of Speech Mixtures

Parallel configuration

Partial separation method(Cascade configuration)

Partial separation method(Cascade configuration)

0510 AM 23

Time domain stage

Frequency domain stage

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 24: Blind Separation of Speech Mixtures

Advantages of Partial Separation methodAdvantages of Partial Separation method

bull Robustness

0510 AM 24

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 25: Blind Separation of Speech Mixtures

Comparison with Adjacent Bands Correlation Method

Comparison with Adjacent Bands Correlation Method

0510 AM 25

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 26: Blind Separation of Speech Mixtures

PS - Partial Separation method with confidence check C1 - Correlation between the adjacent bins without confidence check C2 - Correlation between adjacent bins with confidence check Ha - Correlation between the harmonic components with confidence check PS1 - Partial separation method alone without confidence check

0510 AM 26

Comparison with DOA methodComparison with DOA method

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 27: Blind Separation of Speech Mixtures

My Contribution -IIMy Contribution -II

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

27

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 28: Blind Separation of Speech Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

Underdetermined Blind Source Separation of Instantaneous Mixtures

t

k

0)(0)( 222221 tkStkS

0)(0)( 222221 tkStkS

0)(0)( 112111 tkStkS

0)(0)( 112111 tkStkS

tkX 1

tkX 22x

1x

0510 AM 28

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 29: Blind Separation of Speech Mixtures

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

Mathematical Representation of Instantaneous Mixing

Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773

)(

)(

)(

)( 1

1

1111

tkS

tkS

hh

hh

tkX

tkX

QPQP

Q

P

)(

)(

)(

)( 1

1

1111

ts

ts

hh

hh

tx

tx

QPQP

Q

P

)()()(11

1

1

11

tkS

h

h

tkS

h

h

tkS

h

h

Q

PQ

Q

q

Pq

q

P

)()()()( 21 tkStkStkStk QhhhX

Time domain

Time-Frequency domain

0510 AM 29

P ndash No of mixturesQ ndash No of sources

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 30: Blind Separation of Speech Mixtures

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

00point at 1 Case 11211111 tkStkStk

)()()( 1122111111 tkStkStk hhX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 2222221122 tkStkStk hhX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

0510 AM

0 0

30

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 31: Blind Separation of Speech Mixtures

Q

qqq tkStk

1

hX

2Let Q

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM 31

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 32: Blind Separation of Speech Mixtures

00point at 1 Case 11211111 tkStkStk

)()( 111111 tkStk hX

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()( 222222 tkStk hX

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2

Scalar

ScalarScalar

Scalar

At single source point 1 At single source point 2

111

111

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR 222

222

ofDirection )( ofDirection

ofDirection )( ofDirection

hX

hX

tkI

tkR

Single Source Points in Time-Frequency domainSingle Source Points in

Time-Frequency domain

0510 AM

1111 ( of Direction ( of Direction tkItkR XX 2222 ( of Direction ( of Direction tkItkR XX 32

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 33: Blind Separation of Speech Mixtures

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Perfectly Sparse

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 0

Example

0510 AM 33

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 34: Blind Separation of Speech Mixtures

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

)5()4()3()2()1(

22222

11111

2221

1211

22222

11111

kSkSkSkSkS

kSkSkSkSkS

hh

hh

kXkXkXkXkX

kXkXkXkXkX

21

111 h

hh

22

122 h

hh

0 00 0 00

Example

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse

0510 AM 34

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 35: Blind Separation of Speech Mixtures

Scatter Diagram of the Mixtures when Sources are Sparse

Scatter Diagram of the Mixtures when Sources are Sparse

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

35

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 36: Blind Separation of Speech Mixtures

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

Scatter Diagram of the Mixtures when Sources are Sparse After Clustering

3h

6h

5h

1h4h

2h

0510 AM

No of sources = 6No of mixtures = 2

36

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 37: Blind Separation of Speech Mixtures

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse

3h

6h

5h

1h4h

2h

0510 AM

ObjectiveEstimation of the single source points

No of sources = 6No of mixtures = 2

37

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 38: Blind Separation of Speech Mixtures

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Multi source point

38

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 39: Blind Separation of Speech Mixtures

00point At 1 Case 11211111 tkStkStk

)()(

)()(

111111

111111

tkSItkI

tkSRtkR

hX

hX

Single source point 1

00point At 2 Case 22222122 tkStkStk

)()(

)()(

222222

222222

tkSItkI

tkSRtkR

hX

hX

Single source point 2Scalar

ScalarScalar

Scalar

00point At 3 Case 33233133 tkStkStk

)()()(

)()()(

3322331133

3322331133

tkSItkSItkI

tkSRtkSRtkR

hhX

hhX

)()()( 3322331133 tkStkStk hhX

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

Multi source point

39

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 40: Blind Separation of Speech Mixtures

)(

)(

)(

)( ifonly and if

)( ofDirection )( ofDirection

332

332

331

331

3333

tkSI

tkSR

tkSI

tkSR

tkXItkXR

Average of 15 pairs of speech utterances of length 10 s each

0510 AM

Principle of the Proposed Algorithm for the Detection of Single Source Points

Principle of the Proposed Algorithm for the Detection of Single Source Points

)( ofDirection )( ofDirection tkXItkXR )( ofDirection )( ofDirection tkXItkXR

SSPMSP

40

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 41: Blind Separation of Speech Mixtures

SSPt

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)()(

)()(

112112

111111

tkXjItkXR

tkXjItkXR

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXI

)(

)(

112

111

tkXR

tkXR

)(

)(

112

111

tkXI

tkXIMSP

)( 111 tkX

)( 112 tkX

cos

)()(

)()(

tkXItkXR

tkXItkXR T

Proposed Algorithm for the Detection of Single Source Points

Proposed Algorithm for the Detection of Single Source Points

0510 AM 41

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 42: Blind Separation of Speech Mixtures

Elimination of OutliersElimination of Outliers

SSPs detection

Clus

terin

g

Outlier elimination

0510 AM 42

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 43: Blind Separation of Speech Mixtures

0510 AM

Experimental ResultsExperimental Results

dB6747NMSE

No of mixtures =2 No of sources =6

43

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 44: Blind Separation of Speech Mixtures

Detected Single Source PointsThree mixtures

Detected Single Source PointsThree mixtures

No of mixtures =3 No of sources =60510 AM 44

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 45: Blind Separation of Speech Mixtures

Comparison with Classical Algorithms for Determined Case

Comparison with Classical Algorithms for Determined Case

No of mixtures =2No of sources =2

Average of 500 experimental results

0510 AM 45

-gt

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 46: Blind Separation of Speech Mixtures

Comparison with Method Proposed in [1] Underdetermined case

Comparison with Method Proposed in [1] Underdetermined case

[1] Y Li S Amari A Cichocki D W C Ho and S Xie ldquoUnderdetermined blind source separation based on sparse representationrdquo IEEE Transactions on Signal Processing vol 54 p 423ndash437 Feb 2006

0510 AM

Nor

mal

ized

mea

n sq

uare

err

or (N

MSE

) in

mix

ing

mat

rix e

stim

ation

(dB)

Order of the mixing matrices (PxQ)

46

P ndash No of mixturesQ ndash No of sources

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 47: Blind Separation of Speech Mixtures

Advantages of the Proposed algorithmAdvantages of the Proposed algorithm

Step 1 Convert x in the time domain to the TF domain to get XStep 2 Check the condition

Step 3 If the condition is satisfied then X(k t) is a sample atthe SSP and this sample is kept for mixing matrix estimationotherwise discard the point

Step 4 Repeat Steps 2 to 3 for all the points in the TF planeor until sufficient number of SSPs are obtained

cos

)()(

)()(

tkXItkXR

tkXItkXR T

1) Much simpler constrain the algorithm does not require ldquosingle source zonerdquo

3) The algorithm is extremely simple but effective2) Separation performance is better

0510 AM 47-gt

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 48: Blind Separation of Speech Mixtures

My Contributions ndash III IV and VMy Contributions ndash III IV and V

0510 AM0510 AM

mixtures ge sources mixtures ge sources

mixtures lt sources mixtures lt sources

BSS

Determined Overdetermined

Underdetermined

Instantaneous

Instantaneous

Convolutive

Convolutive

Frequency domain

Frequency domain

Time domain

Time domain

Mixing matrix estimation

Frequency bin-wise separation

Frequency bin-wise separation

Permutation problem

Permutation problem

Source estimation

Automatic detection of no of sources

48

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 49: Blind Separation of Speech Mixtures

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking

Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116

Mask estimation

Mic 1

Mic P

Mixture in TF domain

Separated signals in TF domain

)( tkX P

)(1 tkX

t

t

k

k

)(1 tkY

)(1 tkY

)( tkYQ

)( tkYQ

QP

0510 AM 49

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 50: Blind Separation of Speech Mixtures

Mathematical RepresentationMathematical Representation

Q

q

L

lqpqp Pplnslhnx

1

1

0

1)()()(

)()()()()()()( 11 tkSktkSktkSktk QQqq HHHX

)(

)(

)()(

)()(

)(

)( 1

1

1111

tkS

tkS

kHkH

kHkH

tkX

tkX

QPQP

Q

P

)(

)(

)(

)(

)(

)(

)(

)(

)( 11

1

1

11

tkS

kH

kH

tkS

kH

kH

tkS

kH

kH

Q

PQ

Q

q

Pq

q

P

Time domain

Frequency domain

0510 AM 50

P ndash No of mixturesQ ndash No of sources

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 51: Blind Separation of Speech Mixtures

Single source pointsSingle source pointsInstantaneous mixing

00point at 1 Case 11211111 tkStkStk

)()()(

)()()(

111111

111111

tkSIktkI

tkSRktkR

HX

HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()(

)()()(

222222

222222

tkSIktkI

tkSRktkR

HX

HX

Single source point 2

Real scalar

RealReal

Real scalarReal scalar

Real scalar

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

0510 AM 51

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 52: Blind Separation of Speech Mixtures

Basic Principle of Single Source Points Detection

Basic Principle of Single Source Points Detection

Convolutive mixing

00point at 1 Case 11211111 tkStkStk

)()()( 111111 tkSktk HX

Single source point 1

00point at 2 Case 22222122 tkStkStk

)()()( 222222 tkSktk HX

Single source point 2

Complex scalar

Complex

Complex

Complex scalar

j

H

C

e

cos21

21

UU

UUUUU H

and 20coscos HCH

angle pseudo called is

angleHermitian called is

H The Hermitian angle between the complex

vectors u1 and u2 will remain the same even if the vectors are multiplied by any complex scalars whereas the pseudo angle will change

0510 AM 52

-gt

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 53: Blind Separation of Speech Mixtures

k

Algorithm for Single Source Points Detection

Algorithm for Single Source Points Detection

t

k

tkX 1

tkX 22x

1x

)( 111 tkX

)( 112 tkX

)(

)(

)(111

121

111 tkSkH

kH

11

11

j

jr

)(

)(

112

111

tkX

tkX

11

11

j

jr

)(

)(

112

111

tkX

tkX

)( 111 tkX

)( 112 tkX

)(

)(

)(112

122

112 tkSkH

kH θH2

θH1

θH2

rX

rX

)(

)(cos

11

111

tk

tk H

Hq

0510 AM 53

θH1

OR

1k

1k

1t

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 54: Blind Separation of Speech Mixtures

Clean

Estimated

Mask Estimation by k-means (KM)Mask Estimation by k-means (KM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 54

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 55: Blind Separation of Speech Mixtures

Clean

Estimated

Mask Estimation by Fuzzy c-means (FCM)Mask Estimation by Fuzzy c-means (FCM)

QqttkXtkMtkY pqq 1)()()(

0510 AM 55

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 56: Blind Separation of Speech Mixtures

Automatic Detection of Number of SourcesAutomatic Detection of Number of Sources

0510 AM 56

Cluster validation technique

For c = 2 to cmax

Cluster the data into c clustersCalculate the cluster validation index

EndTake c corresponding to the best cluster as the number of sources

-gt

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points
Page 57: Blind Separation of Speech Mixtures

Elimination of Low Energy PointsElimination of Low Energy Points

0510 AM 57

  • Blind Separation of Speech Mixtures
  • Introduction
  • Slide 3
  • Slide 4
  • Slide 5
  • Approaches for BSS of Speech Signals
  • Slide 7
  • Slide 8
  • Slide 9
  • Slide 10
  • Permutation Problem in Frequency Domain BSS
  • Motivation
  • My Contribution - I
  • Algorithm for Solving the Permutation Problem
  • Existing Method for Solving the Permutation Problem
  • Slide 16
  • Slide 17
  • Slide 18
  • Slide 19
  • Slide 20
  • Proposed algorithm Partial separation method (Parallel configuration) Reference V G Reju S N Koh and I Y Soon ldquoPartial separation method for solving permutation problem in frequency domain blind source separation of speech signalsrdquo Neurocomputing Vol 71 NO 10ndash12 June 2008 pp 2098ndash2112
  • Partial separation method (Parallel configuration)
  • Partial separation method (Cascade configuration)
  • Advantages of Partial Separation method
  • Comparison with Adjacent Bands Correlation Method
  • Comparison with DOA method
  • My Contribution -II
  • Underdetermined Blind Source Separation of Instantaneous Mixtures
  • Mathematical Representation of Instantaneous Mixing Reference V G Reju S N Koh and I Y Soon ldquoAn algorithm for mixing matrix estimation in instantaneous blind source separationrdquo Signal Processing Vol 89 Issue 9 September 2009 pp 1762ndash1773
  • Single Source Points in Time-Frequency domain
  • Slide 31
  • Slide 32
  • Scatter Diagram of the Mixtures When Source are Perfectly Sparse
  • Scatter Diagram of the Mixtures When Source are Not Perfectly Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse
  • Scatter Diagram of the Mixtures when Sources are Sparse After Clustering
  • Scatter Diagram of the Mixtures when Sources are Not Perfectly Sparse
  • Principle of the Proposed Algorithm for the Detection of Single Source Points
  • Slide 39
  • Slide 40
  • Proposed Algorithm for the Detection of Single Source Points
  • Elimination of Outliers
  • Slide 43
  • Detected Single Source Points Three mixtures
  • Comparison with Classical Algorithms for Determined Case
  • Comparison with Method Proposed in [1] Underdetermined case
  • Advantages of the Proposed algorithm
  • My Contributions ndash III IV and V
  • Underdetermined Convolutive Blind Source Separation via Time-Frequency Masking Reference V G Reju S N Koh and I Y Soon ldquoUnderdetermined Convolutive Blind Source Separation via Time- Frequency Maskingrdquo IEEE Transactions on Audio Speech and Language Processing Vol 18 NO 1 Jan 2010 pp 101ndash116
  • Mathematical Representation
  • Single source points
  • Basic Principle of Single Source Points Detection
  • Algorithm for Single Source Points Detection
  • Mask Estimation by k-means (KM)
  • Mask Estimation by Fuzzy c-means (FCM)
  • Automatic Detection of Number of Sources
  • Elimination of Low Energy Points