STQ Workshop, Sophia-Antipolis, February 11th, 2003

Packet loss concealment using audio morphing

Franck Bouteille¹

Pascal Scalart²

Balazs Kövesi²

¹ PRESCOM SA, Lannion, FRANCE

² France Telecom R&D, Lannion, FRANCE

Motivation In packet data networks, excess traffic leads to delays or loss in delivery of

information. In voice communication, long delays are intolerable and network delay budgets have strong influence on the design of packet voice systems.

To increase the tolerance of packet voice systems to lost packets some techniques have been developed.

These techniques do not use the a posteriori information of the next packet that indicates and detects the lost of one or several frames.

However those techniques are not adapted for long lost periods (>15ms) because of the non long-term stationnarity of speech signal.

This a posteriori information is generally available because of the playout buffer management and real time network protocol.

The technique proposed uses the knowledge of the frame received after the last lost one, the models of the last received frames, and a model interpolation to synthesized the missing signal.

Outline

Introduction

Morphing audio principle Voiced / Unvoiced strategy

Modelisation and Interpolation

Blocks concatenation and smoothing

Some results of concealed signal

Comparisons and performances Configuration

Results

Conclusion

Morphing audio principle Context of lost : Previous Frame

Frame A

Missing Signal Next Frame

Frame B

Voiced/Unvoiced strategy

Pitch estimation

Frame A : P0

Pitch estimation

Frame B : P1

UVUV

VUV

UVV

VV

Frame BFrame AP0 , P1P0 , P1 =

P0P0 = P1 , P1Unvoiced

signal

(400 Hz ) 2.5 ms 0, 1 15 ms (67 Hz) P P

When missing signal is defined as unvoiced, Frame A is copied to missing

signal or comfort noise is generated

Morphing audio principle Modelisation and Interpolation:

P0 and P1 are used to estimate the number of necessary intermediate blocks (NbBloc) and the size of these blocks (SizeBloc).

max( 0, 1)SizeBloc P PNbSampleLoss

NbBloc roundSizeBloc

We model the last pitch period vector (X0) of the Frame A (ModP0) and the first pitch period vector (X1) of the Frame B (ModP1). DCT (Dicret Cosinus Transform) is used to model X0 and X1. Resolution is 120 points at 8kHz of sample frequency. Intermediate blocks, , are used in order to transform, in a continuous way, the model vector ModP0 to the model vector ModP1 with linear interpolation of model parameters.

iBlock

IDCT : Inverse Discrete Cosinus Transform.

120

1 00 *

0 1 0 k 120 1

0 1

i

ModP k ModP kBlock n IDCT ModP k i

NbBloc

i NbBloc

n SizeBloc

1

Morphing audio principle Blocks concatenation and smoothing Each block is then copied in the synthesis

frame.

…. ….

Smoothing

Frame A Frame B

Synthesis Frame

0Block 1BlockiBlock 1NbBlocBlock

Smoothing between blocks is realized according to:

(0) (0)* ( 1) (1 (0))* (0)

( 1) : last sample of previous block (or Frame)

y(0) : first sample of current block (or Frame)

( ) ( )* ( 1) (1 ( ))* ( )

1( ) : Smoothing Factor ( ) 1

x x y

x

x i j x i i y i

ii i

NbPSmoothing

1

0 i NbPSmoothing

Morphing audio principle Some results of concealed signal

Conceal frame

Nb sampleNb sample

Original frame

Case of voiced frames of a female speech signal (30ms of missing signal)

Morphing audio principle Some results of concealed signal

Behaviour of the morphing technique during a transition frame (30ms) for male speech signal.

Original frame

Nb sample

Conceal frame

Nb sample

We can notice that the concealed speech to noise transition is more voiced than original frame. In an enhanced morphing technique the voiced duration could be controlled.

Comparisons and performances

Configuration Two speech coders (G.711 and G.723.1) were independently tested, The size frame is 30ms;

Five concealment techniques : Previous Frame Copy: PFC, double Sided Periodic Substitution: DSPS1, ITU-T recommended technique defined for each specific coder: G.711 and G.723.1, GFEC technique2 and Audio Morphing;

Two series of rate were defined: 5 % and 10 %. The losses can appear by burst, but are usually isolated ; The number of sentences was 15 (8 female and 7 male speech files) 1 : J. Tang, "Evaluation of Double Sided Periodic Substitution (DSPS) Method for Recovering Missing

Speech in Packet Voice Communications," IEEE Computers and Communications, pp. 454-458, 1991.2 : B. Kövesi, D. Massaloux, "Method of Packet Errors Cancellation Suitable for any Speech and Sound Compression Scheme", ETSI STQ Workshop, February 2003, Sophia-Antipolis

Ten subjects were participating to an informal test: they were asked to listen to coded speech signals that have been corrected by different concealment techniques

Comparisons and performances Results for G.711 codec

0,00

1,00

2,00

3,00

4,00

5,00

6,00

7,00

Taux 5%

Taux 10%

Rate 5%

Rate 10%

Score (/15)

PFC FECG711 DSPS GFEC MORPHING

Comparisons and performances Results for G.723.1 codec

Note (/15) - G.723.1 - Taux de perte 5% et 10%

0,00

1,00

2,00

3,00

4,00

5,00

6,00

7,00

RTP FECG.723.1 DSPS KB Morphing

Taux 5%

Taux 10%

Rate 5%

Rate 10%

Score (/15)

PFC FECG723 DSPS GFEC MORPHING

Conclusion

Proposed technique improves the quality of the frame correction for strong lost rate (5 % and 10 %);

Morphing audio adds latency (Frame B is required), but is acceptable for application of VoIP;

Another modelisation are possible and voiced condition can be controlled to improve restitution quality