IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM USING DISCRETE MMM WITH CEPSTRAL BASED FEATURES

Transcription

1 4 DAFFODIL INTERNATIONAL UNIVERSITY JOURNAL OF SCIENCE AND TECHNOLOGY, VOLUME 6, ISSUE 2, JULY 20 IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM USING DISCRETE MMM WITH CEPSTRAL BASED FEATURES Md. Rabiul Islam, 2 Md. Fayzur Rahman and 3 Muhammad Abdul Goffar Khan Deparmen of Compuer Science & Engineering 2 Deparmen of Elecrical & Elecronic Engineering 3 Deparmen of Elecrical & Elecronic Engineering Rajshahi Universiy of Engineering & Technology (RUET), Rajshahi-6204, Bangladesh. rabiul_cse@yahoo.com, 2 mfrahman3@yahoo.com, 3 qmag@yahoo.com Absrac: In his paper, an improved sraegy for auomaed ex based speaer idenificaion scheme has been proposed. The idenificaion process incorporaes he Hidden Marov Model echnique. Afer preprocessing he speech, HMM is used in he learning and idenificaion. Feaures are exraced by differen echniques such as RCC, MFCC, MFCC, MFCC, LPC and LPCC which is almos differen in each case. The highes idenificaion rae of 93% has been achieved in he close se ex dependen speaer idenificaion sysem. Keywords: Biomeric Technologies, Auomaic Speaer Idenificaion, Cepsral Coefficiens, Feaure Exracion, Hidden Marov Model.. Inroducion Biomerics is seen by many as a soluion o a lo of user idenificaion and securiy problems now-a- days []. Speaer idenificaion is one of he mos imporan areas where biomeric echniques can be used. There are various echniques o resolve he auomaic speaer idenificaion problem. [2,3,4,5,6,7,8] A wide range of speaer idenificaion applicaions are feasible over dialing-up elephones, including auomaion of operaor assised services, inbound and oubound elemareing, call disribuion by voice, expanded uiliy of a roary phone, reperory dialing and caalog ordering. I is also used in voice conrolled and operaed games and oys, voice recogniion aids for he handicapped and voice conrol of non sraegic funcions in a moving vehicle. The hree imporan echniques for speaer idenificaion are frequenly used. They are he (i) acousic-phoneic approach, (ii) he paern recogniion approach and (iii) he arificial inelligence approach. This paper deals wih he paern recogniion approach. In his wor, close-se ex dependen speaer idenificaion echnique has been considered and Discree Hidden Marov Model has been used as a classificaion echnique. The overall wor has been simulaed using MATLAB based oolbox such as Signal processing Toolbox, Voicebox and HMM Toolbox. 2. Paradigm of Speaer Idenificaion Sysem The basic building blocs of speaer idenificaion sysem are shown in he Figure. The firs sep is he acquisiion of speech from speaers. Then he sar and end poins of speech are deeced. Afer which, pre-emphasis filering echnique has been used. The speech signal is segmened ino overlapping analysis frames. Afer segmenaion, windowing echnique has been applied. Feaures are exraced from he segmened speech. The exraced feaures are hen fed o he DHMM for learning and classificaion. Implemenaion of he speaer idenificaion sysem can be subdivided ino wo pars, (i) he speech signal processing and (ii) he Hidden Marov Model which is used for classificaion. Dae of submission : Dae of accepance :

2 IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM... 5 Acquisiion of Speech Sar and end poin deecion of speech paerns Pre-emphasis filering Speech segmen DAFFODIL INTERNATIONAL UNIVERSITY JOURNAL OF SCIENCE AND TECHNOLOGY, VOLUME 6, ISSUE 2, JULY 20 aion DHMM for Learning Speech feaure exracion Windowing echnique Knowledgebase DHMM for Classificaion Speaer idenificaion Fig. : Bloc Diagram of he proposed auomaed speaer idenificaion sysem 3. Speech signal processing for speaer idenificaion 3.. Acquisiion of Speech Speech acquisiion for his sysem has been done using high qualiy microphone in a sound proof room. To increase he accuracy of his sysem, i is necessary o eep speech acquisiion process noise free. The speech daa are recoded from 20 speaers. The lengh of he speech is abou 3 seconds. Figure 2 shows a sample of he recorded speech. A sampling frequency of 025 Hz wih 6 bis resoluion was used o record he speech voice. The recorded speech was saved in *.wav file forma. 3.2 Sar and End Poin Deecion Speech end poins deecion algorihm has been used o deec he presence of speech, o remove pulse and silences in a bacground noise [9, 0,, 2]. Figure 3 shows he resul afer applying his algorihm over a speech signal. Fig. 2: Sample of he recorded speech signal Fig. 3: Deecion of he necessary speech informaion using sar and end poin deecion algorihm 3.3 Pre-emphasizing Pre-emphasis refers o filering ha emphasizes he higher frequencies. Pre-emphasis has been used o balance he specrum of voiced sounds ha have a seep roll-off in he high frequency region [3, 4, 5]. Figure 3 displays he oupu afer applying he pre-emphasis filering echnique by using he equaion:

3 6 DAFFODIL INTERNATIONAL UNIVERSITY JOURNAL OF SCIENCE AND TECHNOLOGY, VOLUME 6, ISSUE 2, JULY 20 H ( z) = ( az ) () w [ + ] = cos (2Π ), = 0,,..., n (2) n Fig. 4: Speech afer Pre-emphasis filering 3.4 Segmenaion or Frame Blocing In his sep he coninuous speech signal has been bloced ino frames of N samples, wih adjacen frames being separaed by M (M < N). The firs frame consiss of he firs N samples. The second frame begins M samples afer he firs frame and overlaps i by N-M samples. Similarly, he hird frame begins 2M samples afer he firs frame (or M samples afer he second frame) and overlaps i by N-2M samples. This process coninues unil all he speech is accouned. Typically a frame lengh of 0-30 milliseconds is used. A ypical frame overlap is around 25% o 75% of he frame size. The purpose of he overlapping analysis is ha each speech sound of he inpu sequence would be approximaely cenered a some frame [6, 7]. Figure 5 shows a segmened speech signal. 4. Feauer Exracion This sage is very imporan in an ASIS because he qualiy of he speaer modeling and paern maching srongly depends on he qualiy of he feaure exracion mehods. For he proposed ASIS, differen ypes of speech feaure exracion mehods [2, 22, 23, 24, 25, 26] such as RCC, MFCC, MFCC, MFCC, LPC, LPCC have been applied. Figure 6 shows he feaures afer applying differen ypes of feaure exracion echniques. (a) LPC feaures of 2 h order (b) LPCC feaures of 2 h order Fig. 5: Segmened speech 3.5 Windowing In his wor, he purpose of using window is o reduce he effec of he specral arifacs ha resuls from he framing process [8, 9, 20]. From differen ypes of windowing echniques, Hamming window has been chosen for his sysem. The hamming window has been implemened by he equaion [8]: 5. Feaure Condiioning Since DHMM can ae only posiive ineger values as inpu, so i is required o ransform he coninuous valued feaures ino discree valued feaures. I can be performed by using vecor quanizaion mehod. Vecor quanizaion is a sysem for mapping a sequence of coninuous or discree vecors ino a discree codeboo index. The resuls afer applying feaure condiioning are shown in Figure 7.

4 IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM... 7 Fig. 7: Feaure condiioning oupu (c) RCC feaures wih 5 coefficiens (d) MFCC feaures wih 5 coefficiens 6. Speaer modeling For each speaer, an erogodic DHMM (Discree HMM), θ has been buil [27, 28, 29]. The model parameers ( A, B, θ ) have been esimaed o opimize he lielihood of he raining se observaion vecor for he h speaer by using Baum-Welch algorihm. The Baum-Welch re-esimaion formula has been considered as follows [30]: Π i = γ ( i) (3) _ ij a T = = T ξ ( i, j) γ ( i) = T _ γ = ( s,, o = v ) b j ( ) = T = γ ( j) ( j) (4) (5) where, α ( i) aijb j ( o+ ) β + ( j) ξ ( i, j) = N and N _ α ( i) a b ( o+ ) β ( j) N j= i= j= γ ( i) = ξ ( i, j) ij _ j + (e) MFCC feaures wih 5 coefficiens (f) MFCC feaures wih 5 coefficiens Fig. 6: Feaure exracion form of speech In he esing phase, for each unnown speaer o be recognized, he processing shown in Figure 8 has been carried ou. This procedure includes: Measuremen of he observaion sequence O = { o, o2,... on}, via a feaure analysis of he speech corresponding o a speaer.

5 8 DAFFODIL INTERNATIONAL UNIVERSITY JOURNAL OF SCIENCE AND TECHNOLOGY, VOLUME 6, ISSUE 2, JULY 20 Transformaion of he coninuous values of O ino ineger values. Calculaion of model lielihoods for all possible models, P( O θ ), K. Declaraion of he speaer as *speaer whose model lielihood is highes, ha is, * = arg max[ P( O θ ] (6) K In his proposed wor he probabiliy compuaion sep has been performed using he Baum s Forward-Bacward algorihm [30, 3]. 7. Experimenal Resul and Performance Analysis There are some criical parameers (such as frame lengh, frame incremen, number of cepsral coefficiens, number of hidden saes, pre-emphasizing parameer ec) ha affec he performance of DHMM based close-se exdependen speaer idenificaion sysem. The opimal values of he above parameers are chosen o finalize he resul. 7. Experimen on he window shif N In his experimen, he effec of shifing of hamming window has been measured. By seing he window lengh, N L = 5 ms, number of Mel-frequency Cepsral Coefficiens excluding 0 h coefficiens, N MC =5, number of hidden saes, N H =5 and he emphasizing parameer, α = 0.9, we have found he highes speaer idenificaion rae of 85[%] is a 75% window shif as shown in Figure 9. Inpu Speech Preprocessing θ Probabiliy Compuaion HMM for Speaer P ( O θ ) Feaure Exracion 2 θ HMM fir Speaer #2 * = arg max[ P( O θ ] K (c, c 2... c 5 ) Feaure Condiioning Probabiliy Compuaion θ 2 P ( O θ ) HMM for Speaer # Selec Maximum Probabiliy Compuaion n P( O θ ) Fig. 8: Bloc diagram of speaer DHMM recognizer Idenificaion Resul

6 IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM... 9 has been achieved a N H =5 which is shown in Figure. Fig. 9: Performance measuremen according o he window shif 7.2 Experimen on he Pre-emphasized parameer, α In his experimen, he performance of he developed speaer idenificaion sysem has been measured according o he preemphasized parameer α. We have se N L = 5 ms, N =5 ms, N MC =5 and N H =5. We have sudied he value of he parameer ranges from 0.7 o We have found ha he speaer idenificaion performance was 85[%] a α = 0.95 which is shown in Figure 0. Fig. : Resuls afer seing up he hidden saes of DHMM 7.4 Effecs of he window lengh, N L We have chosen he window lengh, N L from 0 ms o 30 ms. By seing N L = 5 ms, N =5 ms, N MC =5 and α = 0.95, he highes performance of 87[%] has been achieved a MFCC based sysem. Figure 2 shows he resul. Fig. 0: Speaer idenificaion rae on he variaion of pre-emphasis parameer 7.3 Experimen of he number of hidden saes of DHMM, N H In he learning phase of DHMM, The hidden saes have been chosen in he range from 5 o 20. We have se N L = 5 ms, N =5 ms, N MC =5, and α = The highes performance Fig. 2: Effec of he window lengh on he idenificaion rae 7.5 Effecs of he number of cepsral coefficiens, N C In his experimen, he number of cepsral coefficiens was varied from 0 o 20 wih a sep size 2. According o he parameers a N L = 5 ms, N =5 ms, N MC =5 and α = 0.95, he highes speaer idenificaion rae was 93[%] which was achieved for MFCC per frame.

7 20 DAFFODIL INTERNATIONAL UNIVERSITY JOURNAL OF SCIENCE AND TECHNOLOGY, VOLUME 6, ISSUE 2, JULY 20 Fig. 3: Speaer idenificaion accuracy according o he number of cepsral coefficiens 8. Conclusion and Observaions The criical parameers such as frame lengh, frame incremen, number of cepsral coefficiens, number of hidden saes and he emphasizing parameer have a grea impac of he idenificaion performance of a DHMM based close se ex dependen ASIS. To find ou he bes performance of his sysem, he opimal values of he above parameers have been seleced effecively. Five experimens have been performed for his purpose. The highes idenificaion rae of 93[%] has been achieved a MFCC. Since he highes speaer idenificaion rae was 93[%], his can saisfy he pracical demand. The performance of his sysem can also be improved by he improvemen of speech signal processing par and by using he hybrid sysem. Open se ex independen speaer idenificaion sysem wih noisy speech can be he furher wor of his sysem. References [] A. Jain, R. Bole, S. Panani BIOMETRICS Personal Idenificaion in Newored Sociey Kluwer Academic Press, Boson, 999. [2] Rabiner, L., and Juang, B.-H., Fundamenals of Speech Recogniion, Prenice Hall, Englewood Cliffs, New Jersey, 993. [3] Jacobsen, J. D., Probabilisic Speech Deecion, Informaics and Mahemaical Modeling, DTU, [4] Jain, A., R.P.W.Duin, and J.Mao., Saisical paern recogniion: a review, IEEE Trans. on Paern Analysis and Machine Inelligence 22 (2000), [5] Davis, S., and Mermelsein, P., Comparison of parameric represenaions for monosyllabic word recogniion in coninuously spoen senences, IEEE 74 Transacions on Acousics, Speech, and Signal Processing (ICASSP), vol. 28, no. 4, pp , Aug [6] Sadaoi Furui, 50 Years of Progress in Speech and Speaer Recogniion Research, ECTI TRANSACTIONS ON COMPUTER AND INFORMATION TECHNOLOGY Vol., No.2, November [7] Locwood, P., Boudy, J., and Blanche, M., Non-linear specral subracion (NSS) and hidden Marov models for robus speech recogniion in car noise environmens, IEEE Inernaional Conference on Acousics, Speech, and Signal Processing (ICASSP), vol., pp , Mar [8] Masui, T., and Furui, S., Comparison of ex-independen speaer recogniion mehods using VQ-disorion and discree/ coninuous HMMs, IEEE Transacions on Speech Audio Process, no. 2, pp , 994. [9] Koji Kiayama, Masaaa Goo, Kaunobu Iou and Tesunori Kobayashi, Speech Sarer: Noise-Robus Endpoin Deecion by Using Filled Pauses, Eurospeech 2003, Geneva, pp [0] S. E. Bou-Ghazale and K. Assaleh, A robus endpoin deecion of speech for noisy environmens wih applicaion o auomaic speech recogniion, in Proc. ICASSP2002, vol. 4, 2002, pp [] A. Marin, D. Charle, and L. Mauuary, Robus speech / non-speech deecion using LDA applied o MFCC, in Proc. ICASSP200}, vol., 200, pp [2] Richard. O. Duda, Peer E. Har, David G. Sro, Paern Classificaion, A Wileyinerscience publicaion, John Wiley & Sons, Inc, Second Ediion, 200. [3] Harringon, J., and Cassidy, S. Techniques in Speech Acousics. Kluwer Academic Publishers, Dordrech, 999. [4] Mahoul, J. Linear predicion: a uorial review. Proceedings of he IEEE 64, 4 (975), [5] Picone, J. Signal modeling echniques in speech recogniion. Proceedings of he IEEE 8, 9 (993),

8 IMPROVEMENT OF THE TEXT DEPENDENT SPEAKER IDENTIFICATION SYSTEM... 2 [6] Clsudio Beccchei and Lucio Prina Ricoi, Speech Recogniion Theory and C++ Implemenaion, John Wiley & Sons. Ld., pp [7] L.P. Cordella, P.Foggia, C. Sansone, and M. Veno, A Real-Time Tex-Independen Speaer Idenificaion Sysem, Proceeding of he 2 h Inern aional Conference on [8] J. R. Deller, J. G. Proais, and J. H. L. Hansen. Discree-Time Processing of Speech Signals. Macmillan, 993. [9] F. Owens. Signal Processing Of Speech. Macmillan New elecronics. Macmillan, 993. [20] F. Harris, On he use of windows for harmonic analysis wih he discree fourier ransform, Proceedings of he IEEE 66, vol. (978), pp.5-84 [2] D. Kewley-Por and Y. Zheng. Audiory models of forman frequency discriminaion for isolaed vowels. Journal of he Acosical Sociey of America, 03(3): , 998. [22] D. O Shaughnessy. Speech Communicaion - Human and Machine. Addison Wesley, 987. [23] E. Zwicer. Subdivision of he audible frequency band ino criical bands (frequenzgruppen). Journal of he Acousical Sociey of America, 33: , 96. [24] S. Davis and P. Mermelsein. Comparison of parameric represenaions for monosyllabic word recogniion in coninuously spoen senences. IEEE Transacions on Acousics Speech and Signal Processing, 28: , Aug 980. [25] S. Furui. Speaer independen isolaed word recogniion using dynamic feaures of he speech specrum. IEEE Transacions on Acousics, Speech and Signal Processing, 34:52 59, Feb 986. [26] S. Furui, Speaer-Dependen-Feaure Exracion, Recogniion and Processing Techniques. Speech Communicaion,Vol. 0, pp , 99. [27] Huang, X. D., Arii, Y., and Jac, M. A., 990. Hidden Marov Models for Speech Recogniion. Edinburgh Universiy Press, Scoland, UK. [28] Hwang, M., and Huang, X., 993. Shared- Disribuion Hidden. Marov Models for Speech Recogniion. IEEE. Trans. on. Speech and Audio Processing, vol., No. 4, pp [29] Baum, L.E., Perie, T., Soules, G., and Weiss, N., 970. A maximizaion echnique occurring in he saisical analysis of probabilisic funcions of Marov chains. The Annals of Mahemaical Saisics, 4, pp [30] Rabiner, L. R., 989. A uorial on hidden Marov models and seleced applicaions in speech recogniion. Proceedings of he IEEE, vol. 77, no. 2, pp [3] Devijver, P. A., 985. Baum's forwardbacward algorihm revisied. Paern Recogniion Leer, 3, pp