ICACSIS 201-1

A Comparison of Backpropagation and LVQ : a casestudy of lung sound recognition

Fadhilah Syafria', Agus Buono'' and Bib Paruhum Silalahi3J: Department of Computer Science and 3Departement of Mathematics

Faculty of Mathematics and Natural Sciences Bogor Agricultural UniversityBogor. Indonesia

Email: ·fadhilahsyafria2102@gmai1.com.2pudesha@yahoo.co.id.3bibparuhuml@yahoo.com

Abstract=Cnie way to evaluate the state of thelungs is by listening to breath sounds usingstethoscope. This technique is knO\\TI as auscultation.This technique is fairly simple and inexpensive, but ithas sorne disadvantage. They are the results ofsubjective analysis. human hearing is less sensitive tolow frequency, em ironmental noise and panern oflung sounds that alrnost similar. Because of thesefactors. misdiagnosis can occur if procedure ofauscultation is not done properly. In this research, willbe made a model of lung sound recognition withneural network approach. Arti ficial neural networkmethod used is Backpropagation (BP) and learningVector Quantization (L VQ). Comparison of these twomethods performed to determine and recommendalgorithms which provide better recognition accuracyof speech recognition in the case of lung sounds. Inaddition to the above two methods. the method of MeiFrequency Cepstrum Coefficient (MFCC) is also usedas method of feature extraction. The results show theaccuracy of using Backpropagation is 93.17%, whilethe value of using the LVQ is R6.R8%. It can beconcluded that the introduction of lung sounds usingBackpropagation method gives better perfonnancecompared to the LVQ method for speech recognitioncases of lung sounds.

I. I;\TRODlT 110;-':

Lung sounds are part of the respiratory sounds.Respiratory sounds include the sound of the mouth andtrachea while lung sounds occur arnund the chest. .Sounds of lungs occurs due to air turbulence when theair enters the respiratory tract during breathing. Thisturbulence occurs because the air flow from the airduct that is wider to narrower airways or vice \ ersa.

ln general. the sound of the lung is divided into two.sounds is the sound of the lungs that are not detectablerespiratory abnormality. whcrcas abnormal lung

sounds are sounds lung disorder.One way to evaluate the state of the lungs is by

listening to breath sounds using a stethoscope. Thistechnique is known as auscultation techniques.Auscultation technique is a basic technique that is usedby physicians to evaluate breath sounds. Thistechnique is fairly simple and inexpensive, but it hasthe disadvantage that the results of a subjectiveanalysis [2]. The results of the analysis of breathsounds using auscultation technique relies on theability. auditory, and experience of the doctor whoperfonned the analysis. In addition, human hearing isless sensitive to low frequency sound is also a problemin th is technique, because the sound of breathingoccupies a low enough frequency. The next problemin auscultation techniques are the problem ofenvironmental noise and the pattern of sounds thatalmost similar between one type of breath sounds andthe other. Because of these factors, misdiagnosis canoccur if the procedure of auscultation is not doneproperly.

Based on the above problems. we interested inconducting research in perfonning speech recognitionof normal lung and lung sounds were detectedinterference (abnormal), Sounds of lungs resulting insome cases of the disease showed a specific patternthat can be recognized. This sound partern can betaken as material for diagnosis [3]. Lung sound will becIassified into four classes, narnely tracheal. vesicular.crackle and wheeze. Tracheal and vcsicular soundindicates nornal lung. whilc the crackle and whcezesound indicates abnormalities in the lungs [4][5].

Voice recognition is an effort in order to the voicecan be recognized or identified so that it can be usedAvoice rccognition system is a computationalapplicarion that is able to identify or verify the soundis autornatically. These systems need to be trained inadvance to be able to recognize the voice and thevoice cIassifies according to the class. For suchpurposcs. an algorithrn is needed to train the speechrecognition system to perfonn more prcc ise .

A system can lcarn and recognize pauerns of sound

ICACSIS 2014

can be done with the application of artificial neuralnerworks. Artificial neural network is a computarionalmethod that mimics the biologicai neural network ofhuman that capable trained to solve problems. Ingeneral, there are several artificial neural networkalgorithms that can be appJied to partern recognition.This research will use two artificial neural networkalgorithms, they are Backpropagation (BP) andLearning Vector Quantization (LVQ). Both methodshave advantages and disadvantages of each.Comparison of these two methods performed todetermine and recommend algorithrns which providebener recognition accuracy of speech recognition inthe case of lung sounds.

ln addition to using an artificial neural networkalgorithm in the classification process, it is no lessimportant in voice recognition is the feature extraction.ln this study, feature extraction methods are used MeiFrequency Cepstrum Coeffisient (MFCC). MFCC is afeature extraction method which gives excellent resultsin classifying normal lung sounds and wheeze sounds[6J.

II. RESEARCH METHOD

Research method conducted is illustrated in FigureI. In the image looks that research begins with aliterature review of the theories necessary for thecompletion of this study, such as lung sound theory.MFCC. Backpropagation. LVQ and MATLABprogranuning.

A. Materials ResearchLung sound data obtained from the repository of

respiratory sounds on the Internet, namely LinmannRepository. The sound data consist of 32 lung sounddata. which is divided into 8 tracheal sounds. 8vesicular sounds. 8 crackle sound and 8 wheeze sound.

Tracheal sound is a sound that is audible at the baseof the neck and larynx. Tracheal sound is very loudand his pitch is high (Fig. 2a). 50 this sound is veryclear sounding compared to another normal lung.Inspiration and expiratory are relatively equal inlength [7J.

Vesicular sounds are normal breath sounds that areheard on the side chest and chest near the stomach.The sounds ware soft with a low pitch (Fig. 2b).Inspiration sounds much stronger than expiratorysounds, Often the expiratory process is hardly audible(7]

Crack/es sounds are a short burst sound that isdiscontinuous (Fig. lc). it is generally more audiblesound m the process of inspiration l,] The conditionsthat is cause the crackles are ARDS, asthma,bronchiectosis. chronic bronchitis. consolidation.carly CflF. nterstitial hlllg disease dan pulnionarvedema [9J.

Start

~~--)t

Uteratcre review

lu", sound theoryMFCC IBackpropagation and LVQ 1MATLAB !

-------,,------'

Data ecquuinon .Lun, sound

Segrnentaucn

Trajnlng Data

Feature ertracuon . ! Feature extr~:

MFCCMfCC

Establlsh the model

(8ackpropagation ~nd lVQ ~---r---~ Model testmg

Companng. Anal!sys. dan

discuanon I. - "r--- J

Documentancn andrepcrts

Figur~ I. Research Method

Wlwe::es sound is a type of sound that IS cominuous,have ii high pitch (Fig. 2d). and more often heard onthe expiratory. This sound occurs when the flow of airthrough the narrowcd airways due to secretions.

foreign body or wounds that preclude [3]. Theduration of the process of inspiration longer than thecxpiratory process. Besides that inspiring soundintensity greater than the expiratory. The conditionsthat is cause the wliee:e are asthma. CIIF, chronichronchitis, COPD dan pulmonarv cdcma [9].

l.~.~~~II ~"t.~I II II III I il Ia. Tracheal .. b.Vesicular .

; I

~II

~d. Crackle d \\'h(,<'7(,

F1sur~ 2 Tvpc of lung '(lund 1<)1

lC ACSIS 2014

B. Segmentation

Lung sound data that has been collected then madethe sound signal segmentation process. This process isdone to cut sound into several small er frames in orderto easily processed. Lung sound signal will be cutbased on one respiratory cycle. Sound signalsegmentation process is done by using the Audacitysoftware. Illustration of signal segmentation processshown in Figure 3. Segmentation process produces 24tracheal sounds. 24\'esicular sounds. 24 crackle soundsand 24 wheeze sounds. 50 the total sounds obtainedfrom the segmentation process is 96 sounds data.

Figure 3. Illustration of sound signal segrnenrauon

C Dala Distrihution (JI Training and Testing Dam

ln general. there are two processes that occur in theartificial neural network algorithm. name ly trainingprocess and testing process. Sound data used in thelearning process is called training data, while thesound data used in the testing process is called testdata. Therefore. the lung sound data that has beensegmented divided into two parts. narnely the trainingdata and test data. The number of lung sound data is96 data. Data lung sounds are then defined as 80training sounds and \ 6 testing sounds data.

D. MFCC Feature ExtractionMFCC feature extraction is a technique uscd to

generate a vector which is used as an identifier [10].The feature is the cepstral coefficient. cepstralcoefficients that are used still considering thepcrception of the human auditory system. M Feetechnique can represent a bener signal than the LPC.LPCC and the other in spcech recognition [II J. This isdue to the workings of the MFCC is based on thefrequency di fTerence that can be captured by thehuman ear so that it can represent how people receivesound signals [12J. Figure 4 is a flow diagram ofMFCC.

(.aH four~r

l,~~"o,,,, (HT)

:t!l.o~tV~· f CC:Pitfum C~fM:l('nh

• 1 DI\(rttf' (O\H'\f' Tf.ln.,form

'I'

Mc:i r r ecve ocvWt.lpp.r\i

hgurc.J D13p-:l111 of :--tFCC

Feature extraction Using MFCC will not rernoveany characteristics or important information of eachlung sound data. In addition, the size of the lung sounddata becomes not too large. Broadly speaking, thereare five stages of MFCC, ie frame blocking,windowing. fast fourier trans form, mei frequencywrapping and cepstrum coefficient.

Lung sound data that has been segmented thenperfonned 'frame blocking' process. Because the sounddata has segmented using audacity software, frameblocking process will read the files that have beenpreviously segmented. Furthermore. windowingprocess using Hamming Window because of simpleformula. The next step is the Fast Fourier Transform(FFT). FFT process is used to convert each frarne thathas been generated from the previous process from thetime domain into the frequency domain, so it can bemore easily observed. Once the signal is convertedfrom the time domain into the frequency domain. thenext step is the mei frequency wrapping process, th isprocess need the filter, th us \"';11 be formed M filterbefore wrapping process is done. Next is the DiscreteCosine Transform (OCT) to get coeffisient cepstrurn.this Coeffisient cepstrum that is an output of theMFCC [13].

E Similaritv Measurement (Classification)Lung sound data that has been obtained their

feature vector through feature extraction process willbe classified using two methods : bacpropagation andLVQ. The results of preliminary processing of data inthe form of voice sarnples that have been segmentedand the extraction process is carried out will be theinput of the two methods.

Broadly spcaking voice recognition lungs out byusing the backpropagation and LVQ has two mainparts. name ly the stages of learning patterns (training)and the stages of pattern recognition ! similaritymeasurement (testing). Learning outcornes datacollected and stored as a learning model that can laterbe used to measure the sirnilarity of a sound signalentering the lungs 50 thai the subsequent lung soundsignais can be recognized. Once the learning phase iscomplete. the next stage is the stage of patternrecognition (similarity measurements).

I) BackpropagationBackpropagation is a ysternatic method of artificial

neural network using supervised learning algorithmand is typically used by the perceptron with rnanylayers (input layer. hidden layer and output layer) tochange the cxisting weights in the hidden layer [\ 4 JBackpropagation is training which type controlledusing weighis adjustment pancrns to achieve theminimum value of the error between the output of theprcdiction rcsults with real output. To gct a networkerror. the forward propagation phase (fcedforward)must be done before.

lC ACSIS 201-1

In the feedforward phase, each unit receives inputsignais and forwards the signal to all units in the layerabove it (the hidden layer). Each unit in the hiddenlayer summing the input signais that is receives. In thisstep activation function is used to calculate the outputsignal to be sent to all units on it (the output layer).This step done as much as the number of hiddenlayers. Then, for each unit of output summing theoutput signais. In this step activation function is alsoused to caJculate the output signal [15].

After the feedforward phase. the next step isbackforward phase. In this phase, eaeh output unitreceives a target pattern associated with the traininginput pattern and then caJculate his error information.Then calculate the correction weights and biascorreetion. This step done as much as the number ofhidden layers. Each hidden unit delta sumrning inputs(from units located in the layer above it). Multiply thedelta value by the derivative of activation function tocaJculate error information, and then caJculate theeorrection weights and bias correction. The final step,for each unit of output fixing weights and bias, as wellas the hidden units are also fixing weights and bias[ 15].

2) Learning Vector Quantization (L VQ)LYQ algorithm is a method for training cornpetitive

layers of the supervised [16]. Competitive layer willautornatically learn to cJassify the given input vector,The approach taken is by cJassify the input veetorbased on proximity distanee of the input vector toagent vector (eucJidean distanee method).

LYQ network structure is a two-layer neuralnetwork is cornposed of input layer and output layer.Input layer contains neurons as ma ny as dimensionalinput, output layer contains neurons as many as thenumber of cJasses [17]. These two layers areconnected by link between each neuron that has acertain weight is called the agent vector.

LYQ algorithm is a method of artificial neuralnetwork based on competition with a mechanismsquared euelidean distanee in picking winners agentvector to detennine the category of the input vector.ln LYQ network. the learning proeess undertaken are asupervised learning. where providing input areaccompanied by the expecied output information.Network always directed to determine the mostappropriate output unit with the target of input vectorsthrough shi ft position of agent vector. If the vectors oftraining data are grouped together with agent vectorwinner, then the agent vector is shifted come near tovector of training data by the equation :

lVi~(n,,"') = I\',~(old) + aC.\': - 11'/0(0) (1)

If \ ectors the training data are not grouped togetherwith the agent \ cc tor winner, then the agent vectorshifted away from the training vector is expressed bythc equation :

1V,~(l1e\l')= 1V,~(old)- a(.\': -lVj:(old) (2)

\\'here :

a Learning rateJV Weight of agent vectorX Weight ofinput vector

111. RESULTS A.'1D DISCUSSION

A. Testing Results of LI'Q Method

Testing at this st age to look at the level ofrecognition accuracy using LYQ. Testing is done bychanging the number of eoeffieients MFCC and thevalue of learning rate to see its effect on the level ofrecognition accuracy. The number of MFCCeoeffieients used are tv1FCC 13. MFCC IS. MFCC 20dan MFCC 3D, where as the learning rate us ed are 0,1.0,2.0,3, DA dan 0,5.

LYQ implernentation IS done 111 Matlabenvironment with 1vqne r, function, Learningfunctions used are LVQ ') that on Mat labirnplemented with funetions 1ea rn 1vq2 .

According to experirnent on LYQ by varying theparameter values the learning rate and coefficientsMFCC, obtained accuracy levels vary in testing withi6i Jala as presented in Table I, There was no generalpattern of the relationship betwecn MFCC coefficientand learning rate, but bath pararneters affect theaccuracy freely. However, the value of the highestaccuracy in LVQ obtained with 20 MFCC coefficientsand the of 0,2 learning rate with an accuracy 97,02°0(marked in bold).

T..\13U: I:\en'RA( Y OF T1SII:--(, 1:--L \'Q

1\11'('(' Learning Rate

( 'ocfficicnt-,0,1 0, 2 O.. 1 (J .j 0, 5

11 9tl I X 9.~...l5 lI-10X X2-1-1 7(, I 'I

15 RII,99 'I} -15 9137 Rl.R5 X2 1-1

20 XlI.5'l 97.02 76.19 X-1.R2 R6 OI

30 92,26 XC"OI X~.R2 90A 7 /7 O~

Overall. testing with LVQ can achieve the highestaccuracy of 100° 0, l Iowever. in a minority ofcxperirnenrs obtained an accuracy of 0°0. Ie no data iscorrecrly identificd. This IS due to the perforrnancc ofthe L\'Q trauung up and down during the trainingproccss iteration or not towards convergcnt. Forcxarnple. tt was detcrmincd that the rnaximum error of0.0 I. howcvcr when the training proeess untiI themaximurn iteration nev er bcen obtained the error rate.\ en the error becomes lughcr as the iterauon gocsthough the prcx iOl!> nerauon error of low \ alue

B, Testing Rcsults (JIBackpropagation .I/crhot/

Testing at th is sragc to look at thc level ofrccognition accuracy using Backprnpagation. Testingi, done by ch.mg ing the number of coefficients \IFCCand the value of learning rate to see ih effcct on the

lf)( )

lC ACSIS 2014

level of recognition accuracy. The number of MFCCcoefficients used are MFCC 13. MFCC 15. MFCC 20dan MFCC 30. whereas the learning rate us ed are 0.1,0.2.0.3,0.4 dan 0.5.

Backpropagation implementation is done in Matlabenvironment with newff Activation function used issigmoid (log 5 i g) in the hidden layer and a linearfunction (pu r e 1i n) in the output layer. Trainingalgorithm used is gradient descent (t rai ngd).

According to experiment on Backpropagation byvarying the parameter values the learning rate andcoefficients MFCC. obtained accuracy levels vary intesting with test data as presented in Table 2. Just likethe LVQ. not obtained a clear general pattern of therelationship berween MFCC coefficient and learningrate. However, from the table it can be seen that theaccuracy tends to increase with increasing learningrate with an increase in the flucruating. In some cases,the value of learning rate that is too high can make alower accuracy.

the value of the highest accuracy lnBackpropagation obtained with 13 MFCC coefficientsand the of 0.-1 learning rate with an accuracy 97.32°0(Table 2). Backpropagation method also can achievethe highest accuracy of 100% as in LVQ, but in sorneexperiments, the accuracy is quite low on a particularclass can even reach 0%.

TABLE IIACCl'RACY or TESTII\"C; II\" BACKPROPAC;,\TIOI\"

fo-IFCC LcarnmgRate

CocfficicntsO. 1 O. (J 3 O.~ 0 5

U 91 % 96 -13 <)-11>3 97..'2 96 73

15 );988 94.94 9H)5 95.l'\3 9643

20.· 87.5 91 91> 93.75 '15.54 9-1.35

30 80.36 89.5R 95.24 91.96 94.9-1

C. Effect ofXutnbc: o{,\fFCC Coefficient AgainstAccllracl"

From the cntire experiment. the average accuracyobtained at each value of the coefficient is shown inFigure 5. From the graph it can be seen that there is atendency that the higher the coefficient MFCC makeslower accuracy 111 LVQ and Backpropagation.although the decrease was not significant. Accuracy atMFCC coefficient 13 tend to be higher th an theaccuracy of the MFCC coefficient greater. This is dueto the higher coefficient of MFCC impact on thedimensions of a larger data. Largc data dirncnsionsmake the generalization capability of A\,1\,1 is lowcr sothe accuracy is dccrcased.

D. E[{cct o] Learning Rate Against Accllrac\'

lnfluence of the learning rate throughout thecxpcrirnent to the average accuracy is shown in Figuren. From the graph. in LVQ there is a tendcncy that

highest accuracy occur on the learning rate is not toolow or too high. If the learning rate is 100 high,accuracy will be decrease because of the LVQ that dothe change of network weights with the nature of thecompetition, ie, increase or decrease the weight of thenetwork. so that the learning rate is too large causingdrastic changes at each iteration 50 that the weight ofthe network becomes unstable. However, if thelearning rate is too low. the learning process will beIon.

Graph of accuracv Based on MfCC coeffident

10098

z 9694~ 91e 90e 88c 86.. 84S180

13

elVQ(Tr.ini,,€ 0"., 88.31-~.-8lVQ(T es ling 0..1.) 87.26-- ----

• BP (Tr .lining 0')1.)) 96.58

• BP rre~ti~ D~td) 95,47

-'

15

87,81

87.56

95,94

94.13

30

87,11

85.95

91.59

90.42

Figure 5 Graph of accuracy Based on MFCC coefficient

~------- ~~~~~~~~~-Graph of eccuracv Based on terning Rote

100989694n908lj8684H780

0.1 0.1 0.3 0.4 O.SalVQ(iI.llnu,,€ OJI,)} 91.37 93.5' B7.37 85.46 BI.4Q

• l IJU (h'\hng Odld) 90.}~ 97.48 86.lR 84.B9 8O.l6

• ijP (f',lning 0.11.)) 89.91 93.~3 9~.~ 90.64 9/.Jl

• BV (! e~tl~ OaL)) 8/.41 93.1 J 94.42 9~.16 ss.s:----.~ - - ~ -Figure I>Graph of accuracy Based on l.enrning' Rate

ln contrast. in Backpropagation, increased learningrate rends to increase the accuracy. This is due to thealgorithm Backpropagation just changing the networkweights by summing up the weighis (without the

nature of the competition) so that the network weightrernains stable.

E. lntroductorv Rate of Each LIII/g Sounds

Average of accuracy of the entire experiment basedon the type of lung sounds is shown in Figure 7. Fromthe graph it can be seen that the type of tracheal soundhas the highest accuracy. rncaning that most can bedistinguished from other types of sound. [t is. asmentioned previously (Figure 7) that a tracheal soundsignals has the form the most different than the othcrthrce types of sound. Meanwhile, the sound ofvesicular type. crackle. and wheezc has littleresemblance so that iIS accuracy is lower. meaning thatthe sound of vesiculer is sornetimes identified as acrackle type or whecze and vice versa,

to I IS13:\ !J7x-1)7~)-1121-22.-)

ICACSIS 201-1

Graph of accuracy of eadi type sound

100 ---~~-------98969492908886848280

• LVQ(Tr~jnil"€ D.lt.l) 94,6 85 87.7 84

al VO (To'ting Data) 94,6 81.3 87,8 84---------_._- - ----~--• BP (Tr oiining D,H.l) 99,4 93,3 91 94,7

a SP ITo,ting Data) 98,2 91,4 88,6 94,4

Figure 7 Graph of accuracy in Each Type Sound

F. Comparison of Accuracy Rate Backpropagationand L VQ MethodComparison of the accuracy of LYQ and

Backpropagation method in identifying lung sounds ispresented in Figure 8, Overall, backpropagationmethod gives better accuracy results than LYQmethod, with an overall average accuracy of 94,58%for training data and 93,17% for test data, while LVQonly offer results of R7,83% for training data and86,88% for test data, Testing the training data providegreater accuracy results because the data is also usedin the training process, while the test data were notincluded during training,r---

Comparison of the accuracy of lVQ method and BP

-~-~--,

1009896

g 9'

~ 92

e 90e AA;: 86

84

8}

80Ir ainmg O.lh

_tIJQ 8/,877,)

• BP 94,5775

T(>$tng O.t.1

86,8/')

93.17l5

Figure 8 Comparison of Accuracy graphs of LVQ andBackpropagation

Although LYQ method gives lower accuracy, butthe method is simpler than computing aspects 50 thattraining time is faster than the Backpropagationmethod, This is consistent with studies conductedHawickhorst ef al. (1995) that Backpropagation givesberter accuracy results than LYQ. but highercomputational complexity [18],

tv. CO~CLL:SIO~

From the research th at has been done. it can beconcluded that it has made a model of lung soundrecogniuon with MFCC feature extraction.Classification with LVQ and Backpropagation methodcan be applied for the idenrification of lung sounds. byadjust paramcrers MFCC coefficient and learning ratethat maximizes the accuracy, Backpropagation methodhas bener accuracy than LVQ with all ayeragcaccuracy on the test data was 93.17n o. whilc Ro.RRO n

of LYQ, In addition, tracheal sound can be recognizedbetter than the lung sound types vesicular, Crackle,and wheeze.

REFERENCES

[t] A. Cohen. D, Landsberg, "Analysis and automaticclassification of breath sounds", Biomedicai Engineering.IEEE Transactions on. (9). 585-590. 1984,

[2] H. Kiyokawa, MDM. Greenberg. K. Shirota. H, Pasterkamp,"Auditory Detection of Simulated Crackels in BreathSounds", CHEST. t 19(6): 1886-1892.2013

[3] A. Rizal. MD. Samudra, I. lwut. V. Suryani. "PengenalanSuara Paru Menggunakan Spektogram dan K-MeanClusrering", Proceeding SI TIA20 I O. Februari 20 I0

[4] AA, Abaza. 18. Day. JS. Reynolds. AM. Mahmoud. WTGoldsmith, WG, McKinncy. EL Petsonk. DG. Frazer,"Classification of voluntary cough sound and airflow patternsfor detecting abnormal pulrnonary function". Cough. 5(8).2009,

[5] A. Gurung. CG. Scrafford. JM. Tielsch. OS. Levine. w.Checkley \Ii, "Computerized lung sound analysis asdiagnostic aid for the detection of abnormal lung sounds: asysternatic review and meta-analysis". Respir Med, 105(9):1396-1403,20 II.

f6] M. Bahoura, "Pauern Recognition Methods Applied toRespiratory Sounds Classification into Normal and WheezeClasses. Computcrs and Biolog- and ~kdiClne. 39(9):824-8~3, 2009.

[71 A. Rizal. L Anggraeni, V Suryani, "Pengenalan Suara Paru-Paru Normal Menggunakan LPC dan Jaringan Syaraf TiruanBack-Propagatiou'. Proceeding EECClS2006, Mei 2006.

[81 T. Katila. P Piirila. K. Kallio. E, Paajanen. T. Rosqvist. AR,Sovijarvi, "Original waveform of lung sound crackles: a casestudy of the effect of high-pass filtration" Journal of AppliedPhvsiologv. 7/(6),2173-2177, t991.

191 ~lZ. Ramadhan, "Perancangan Sistem Instrumentasi untukIdentifikasi dan Analisis Suara Paru-Paru Menggunakan DSPH1S320CMI6T [Essay]", lkpok[IDI Universitaslndonesia, 2012.

1101 K. Patcl, KK. Prasad. "Spccch Rccognition and VcrificationUsing MFCC & "Q". Intcrntional Journal. 20) 3,

[111 A. Buono. "Representasi nilai hos dan model MFCC sebagaiekstraksi ciri pada sistem indentifikasi pembicara dilingkungan ber-noiscmcnggunakan HMM [disscrtation [".Depok: Program Studi llmu Komputer. Universitas lndonesia .2009.

f 121 L. Muda, M, Begam. L Elamvazuthi. "Voice RccogniiionAlgoriihms Using Mei Frcquencv Ccpstral CocfficicntIMFCC) and Dynamic Tune Warping IDTW) Tcchniqucs",Journal of Cornpuring. Volume 2, lssuc J, 2010.

[131:--'1. Slancy. "Auditory Toolbox Interval ResearchCorporation". Tech. Rcp, ID, 1998,

r 141 F. Suhandi, "Prediksi Harga Saham dcngan PcndckatanAnificial Ncural 1\elwork mcnacunakan AlgoritmaBackpropagarion", 2009.

[151 L. Fausett. "Fundamcnrals of ncural nctworks architccturesalgorithms and applications". Florida l.atitudc of Technologv,199~.

11(>1 S, Kusumadewi. S. Hartati. "lntcgras: SIStem Fuzzy danJaringan Syaraf', Yogyakana. (iraha llmu, ~O IO.

1171 L, Rahadianu. "Pengembangan Algoritma PembelajaranBerbasis Dimensi serta Kornparasinva rcrhadao PembelajaranBerbasiskan Vektor rada FII::\ -Xcuro Learning'·l'dorQIIOllfi:Ol/o1l untuk Pengenalan Citra \\-3)3h Frontal!Essa\ [". Depok [IDI Fa-rlkorn l.nivcrsuas lndoncsra. ~009

[I XI BA Hawickhorst. S:\ Zahonan. K Rajagopal. ..:\Comparison of Thrcc Ncural :--'-ctwork Architccrurcs forAuiornauc Specch Rccognition" DI dalam Dagli CH, el al.editor Procccdmgs of Ihe Aruficiol Xcurol Xctw orks in1:-"gmC(,f'lng . .-I.\"SII:"'95. ~'cw York :\S\1E Press. hlm ~~1·"

[191 ,1r

lIl2