Improvement of Text Dependent Speaker Identification System Using Neuro-Genetic Hybrid Algorithm in Office Environmental Conditions

Transcription

1 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, 2009 ISSN (Onlne): ISSN (Prnt): Improvement of Text Dependent Speaker Identfcaton System Usng Neuro-Genetc Hybrd Algorthm n Offce Envronmental Condtons Md. Rabul Islam 1 and Md. Fayzur Rahman 2 1 Department of Computer Scence & Engneerng Rajshah Unversty of Engneerng & Technology (RUET), Rajshah-6204, Bangladesh 2 Department of Electrcal & Electronc Engneerng Rajshah Unversty of Engneerng & Technology (RUET), Rajshah-6204, Bangladesh Abstract In ths paper, an mproved strategy for automated text dependent speaker dentfcaton system has been proposed n nosy envronment. The dentfcaton process ncorporates the Neuro- Genetc hybrd algorthm wth cepstral based features. To remove the background nose from the source utterance, wener flter has been used. Dfferent speech pre-processng technques such as start-end pont detecton algorthm, pre-emphass flterng, frame blockng and wndowng have been used to process the speech utterances. RCC, MFCC, MFCC, MFCC, LPC and LPCC have been used to extract the features. After feature extracton of the speech, Neuro-Genetc hybrd algorthm has been used n the learnng and dentfcaton purposes. Features are extracted by usng dfferent technques to optmze the performance of the dentfcaton. Accordng to the VALID speech database, the hghest speaker dentfcaton rate of % for studo envronment and % for offce envronmental condtons have been acheved n the close set text dependent speaker dentfcaton system. Key words: Bo-nformatcs, Robust Speaker Identfcaton, Speech Sgnal Pre-processng, Neuro-Genetc Hybrd Algorthm. 1. Introducton Bometrcs are seen by many researchers as a soluton to a lot of user dentfcaton and securty problems now a days [1]. Speaker dentfcaton s one of the most mportant areas where bometrc technques can be used. There are varous technques to resolve the automatc speaker dentfcaton problem [2, 3, 4, 5, 6, 7, 8]. Most publshed works n the areas of speech recognton and speaker recognton focus on speech under the noseless envronments and few publshed works focus on speech under nosy condtons [9, 10, 11, 12]. In some research work, dfferent talkng styles were used to smulate the speech produced under real stressful talkng condtons [13, 14, 15]. Learnng systems n speaker dentfcaton that employ hybrd strateges can potentally offer sgnfcant advantages over sngle-strategy systems. In ths proposed system, Neuro-Genetc Hybrd algorthm wth cepstral based features has been used to mprove the performance of the text dependent speaker dentfcaton system under nosy envronment. To extract the features from the speech, dfferent types of feature extracton technque such as RCC, MFCC, MFCC, MFCC, LPC and LPCC have been used to acheve good result. Some of the tasks of ths work have been smulated usng Matlab based toolbox such as Sgnal processng Toolbox, Vocebox and HMM Toolbox. 2. Paradgm of the Proposed Speaker Identfcaton System The basc buldng blocks of speaker dentfcaton system are shown n the Fg.1. The frst step s the acquston of speech utterances from speakers. To remove the background noses from the orgnal speech, wener flter has been used. Then the start and end ponts detecton algorthm has been used to detect the start and end ponts from each speech utterance. After whch the unnecessary parts have been removed. Pre-emphass flterng technque has been used as a nose reducton technque to ncrease the ampltude of the nput sgnal at frequences where sgnal-to-nose rato (SNR) s low. The speech sgnal s segmented nto overlappng frames. The purpose of the overlappng analyss s that each speech sound of the nput

2 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, sequence would be approxmately centered at some frame. After segmentaton, wndowng technque has been used. Features were extracted from the segmented speech. The extracted features were then fed to the Neuro-Genetc hybrd technques for learnng and classfcaton. From dfferent types of wndowng technques, Hammng wndow has been used for ths system. The purpose of usng wndowng s to reduce the effect of the spectral artfacts that results from the framng process [30, 31, 32]. The hammng wndow can be defned as follows [33]: Πn N N cos, ( ) n ( ) w ( n) = (3) N 2 2 0, Otherwse 4. Speech parameterzaton Technques for Speaker Identfcaton Fg. 1 Block dagram of the proposed automated speaker dentfcaton system. 3. Speech Sgnal Pre-processng for Speaker Identfcaton To capture the speech sgnal, samplng frequency of Hz, samplng resoluton of 16-bts, mono recordng channel and Recorded fle format = *.wav have been consdered. The speech preprocessng part has a vtal role for the effcency of learnng. After acquston of speech utterances, wener flter has been used to remove the background nose from the orgnal speech utterances [16, 17, 18]. Speech end ponts detecton and slence part removal algorthm has been used to detect the presence of speech and to remove pulse and slences n a background nose [19, 20, 21, 22, 23]. To detect word boundary, the frame energy s computed usng the sort-term log energy equaton [24], n + N 1 t = n 2 E = 10 log S ( t ) (1) Pre-emphass has been used to balance the spectrum of voced sounds that have a steep roll-off n the hgh frequency regon [25, 26, 27]. The transfer functon of the FIR flter n the z-doman s [26] Ths stage s very mportant n an ASIS because the qualty of the speaker modelng and pattern matchng strongly depends on the qualty of the feature extracton methods. For the proposed ASIS, dfferent types of speech feature extracton methods [34, 35, 36, 37, 38, 39] such as RCC, MFCC, MFCC, MFCC, LPC, LPCC have been appled. 5. Tranng and Testng Model for Speaker Identfcaton Fg.2 shows the workng process of neuro-genetc hybrd system [40, 41, 42]. The structure of the multlayer neural network does not matter for the GA as long as the BPNs parameters are mapped correctly to the genes of the chromosome the GA s optmzng. Bascally, each gene represents the value of a certan weght n the BPN and the chromosome s a vector that contans these values such that each weght corresponds to a fxed poston n the vector as shown n Fg.2. The ftness functon can be assgned from the dentfcaton error of the BPN for the set of pctures used for tranng. The GA searches for parameter values that mnmze the ftness functon, thus the dentfcaton error of the BPN s reduced and the dentfcaton rate s maxmzed [43]. 1 H ( Z ) = 1 α. z, 0 α 1 (2) Where α s the pre-emphass parameter. Frame blockng has been performed wth an overlappng of 25[%] to 75[%] of the frame sze. Typcally a frame length of mllseconds has been used. The purpose of the overlappng analyss s that each speech sound of the nput sequence would be approxmately centered at some frame [28, 29].

3 44 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, 2009 Fg.2 Learnng and recognton model for the Neuro-Genetc hybrd system. The algorthm for the Neuro-Genetc based weght determnaton and Ftness Functon [44] s as follows: Algorthm for Neuro-Genetc Weght determnaton: { 0; Generate the ntal populaton P of real coded chromosomes C j each representng a weght set for the BPN; Generate ftness values F j for each C j P usng the algorthm FITGEN(); Whle the current populaton P has not converged { Usng the cross over mechansm reproduced offsprng from the parent chromosome and performs mutaton on offsprng; +1; Call the current populaton P ; Calculate ftness values F j for each C j P usng the algorthm FITGEN(); Extract weght from P to be used by the BPN; Algorthm for FITGEN(): {Let ( I, T j ), =1,2,..N where I = ( I 1, I 2, I l ) and T = ( T 1, T 2, T l ) represent the nput-output pars of the problem to be solved by BPN wth a confguraton l-m-n. { Extract weghts W from C ; Keepng W as a fxed weght settng, tran the BPN for the N nput nstances (Pattern); Calculate error E for each of the nput nstances usng the formula: E, = 2 ( T j O j ) (3) j Where O s the output vector calculated by BPN; Fnd the root mean square E of the errors E, I = 1,2,.N.e. E = E N (4) Now the ftness value F for each of the ndvdual strng of the populaton as F = E; Output F for each C, = 1,2,.P; 6. Optmum parameter Selecton for the BPN and GA 6.1 Parameter Selecton on the BPN There are some crtcal parameters n Neuro-Genetc hybrd system (such as n BPN, gan term, speed factor, number of hdden layer nodes and n GA, crossover rate and the number of generaton) that affect the performance of the proposed system. A trade off s made to explore the optmal values of the above parameters and experments are performed usng those parameters. The optmal values of the above parameters are chosen carefully and fnally fnd out the dentfcaton rate Experment on the Gan Term, η In BPN, durng the tranng sesson when the gan term was set as: η 1 = η 2 = 0.4, spread factor was set as k 1 = k 2 = 0.20 and tolerable error rate was fxed to 0.001[%] then the hghest dentfcaton rate of 91[%] has been acheved whch s shown n Fg.3. Fg. 3 Performance measurement accordng to gan term Experment on the Speed Factor, k The performance of the BPN system has been measured accordng to the speed factor, k. We set η 1 = η 2 = 0.4 and tolerable error rate was fxed to 0.001[%]. We have studed the value of the parameter rangng from 0.1 to 0.5. We have found that the hghest recognton rate was 93[%] at k 1 = k 2 = 0.15 whch s shown n Fg.4.

4 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, Fg. 4 Performance measurement accordng to varous speed factor Experment on the Number of Nodes n Hdden Layer, N H In the learnng phase of BPN, We have chosen the hdden layer nodes n the range from 5 to 40. We set η 1 = η 2 = 0.4, k 1 = k 2 = 0.15 and tolerable error rate was fxed to 0.001[%]. The hghest recognton rate of 94[%] has been acheved at N H = 30 whch s shown n Fg.5. Fg. 6 Performance measurement accordng to the crossover rate Experment on the Crossover Rate Dfferent values of the number of generatons have been tested for achevng the optmum number of generatons. The test results are shown n the Fg.7. The maxmum dentfcaton rate of 95[%] has been found at the number of generatons 15. Fg. 5 Results after settng up the number of nternal nodes n BPN. Fg.7 Accuracy measurement accordng to the no. of generatons. 6.2 Parameter Selecton on the GA To measure the optmum value, dfferent parameters of the genetc algorthm were also changed to fnd the best matchng parameters. The results of the experments are shown below Experment on the Crossover Rate In ths experment, crossover rate has been changed n varous ways such as 1, 2, 5, 7, 8, 10. The hghest speaker dentfcaton rate of 93[%] was found at crossover pont 5 whch s shown n the Fg Performance Measurement of the Text- Dependent Speaker Identfcaton System VALID speech database [45] has been used to measure the performance of the proposed hybrd system. In learnng phase, studo recordng speech utterances ware used to make reference models and n testng phase, speech utterances recorded n four dfferent offce condtons were used to measure the accurate performance of the proposed Neuro-Genetc hybrd system. Performance of the proposed system were measured accordng to varous cepstral based features such as LPC, LPCC, RCC, MFCC, MFCC and MFCC whch are shown n the followng table. Table 1: Speaker dentfcaton rate (%) for VALID speech corpus Type of MFCC envronments MFCC MFCC RCC LPCC Clean speech utterances Offce envronments

5 46 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, 2009 speech utterances Table 1 shows the overall average speaker dentfcaton rate for VALID speech corpus. From the table t s easy to compare the performance among MFCC, MFCC, MFCC, RCC and LPCC methods for Neuro-Genetc hybrd algorthm based text-dependent speaker dentfcaton system. It has been shown that n clean speech envronment the performance s [%] for MFCC, MFCC and LPCC and the hghest dentfcaton rate (.e [%]) has been acheved at MFCC for four dfferent offce envronments. 8. Concluson and Observatons The expermental results show the versatlty of the Neuro- Genetc hybrd algorthm based text-dependent speaker dentfcaton system. The crtcal parameters such as gan term, speed factor, number of hdden layer nodes, crossover rate and the number of generatons have a great mpact on the recognton performance of the proposed system. The optmum values of the above parameters have been selected effectvely to fnd out the best performance. The hghest recognton rate of BPN and GA have been acheved to be 94[%] and 95[%] respectvely. Accordng to VALID speech database, 100[%] dentfcaton rate n clean envronment and [%] n offce envronment condtons have been acheved n Neuro-Genetc hybrd system. Therefore, ths proposed system can be used n varous securty and access control purposes. Fnally the performance of ths proposed system can be populated accordng to the largest speech recognton database. References [1] A. Jan, R. Bole, S. Pankant, BIOMETRICS Personal Identfcaton n Networked Socety, Kluwer Academc Press, Boston, [2] Rabner, L., and Juang, B.-H., Fundamentals of Speech Recognton, Prentce Hall, Englewood Clffs, New Jersey, [3] Jacobsen, J. D., Probablstc Speech Detecton, Informatcs and Mathematcal Modelng, DTU, [4] Jan, A., R.P.W.Dun, and J.Mao., Statstcal pattern recognton: a revew, IEEE Trans. on Pattern Analyss and Machne Intellgence 22 (2000), pp [5] Davs, S., and Mermelsten, P., Comparson of parametrc representatons for monosyllabc word recognton n contnuously spoken sentences, IEEE 74 Transactons on Acoustcs, Speech, and Sgnal Processng (ICASSP), Vol. 28, No. 4, 1980, pp [6] Sadaok Furu, 50 Years of Progress n Speech and Speaker Recognton Research, ECTI TRANSACTIONS ON COMPUTER AND INFORMATION TECHNOLOGY, Vol.1, No.2, [7] Lockwood, P., Boudy, J., and Blanchet, M., Non-lnear spectral subtracton (NSS) and hdden Markov models for robust speech recognton n car nose envronments, IEEE Internatonal Conference on Acoustcs, Speech, and Sgnal Processng (ICASSP), 1992, Vol. 1, pp [8] Matsu, T., and Furu, S., Comparson of text-ndependent speaker recognton methods usng VQ-dstorton and dscrete/ contnuous HMMs, IEEE Transactons on Speech Audo Process, No. 2, 1994, pp [9] Reynolds, D.A., Expermental evaluaton of features for robust speaker dentfcaton, IEEE Transactons on SAP, Vol. 2, 1994, pp [10] Sharma, S., Ells, D., Kajarekar, S., Jan, P. & Hermansky, H., Feature extracton usng non-lnear transformaton for robust speech recognton on the Aurora database., n Proc. ICASSP2000, [11] Wu, D., Morrs, A.C. & Koreman, J., MLP Internal Representaton as Dscmnant Features for Improved Speaker Recognton, n Proc. NOLISP2005, Barcelona, Span, 2005, pp [12] Kong, Y., Heck, L., Wentraub, M. & Sonmez, K., Nonlnear dscrmnant feature extracton for robust textndependent speaker recognton, n Proc. RLA2C, ESCA workshop on Speaker Recognton and ts Commercal and Forensc Applcatons, 1998, pp [13] Ismal Shahn, Improvng Speaker Identfcaton Performance Under the Shouted Talkng Condton Usng the Second-Order Hdden Markov Models, EURASIP Journal on Appled Sgnal Processng 2005:4, pp [14] S. E. Bou-Ghazale and J. H. L. Hansen, A comparatve study of tradtonal and newly proposed features for recognton of speech under stress, IEEE Trans. Speech, and Audo Processng, Vol. 8, No. 4, 2000, pp [15] G. Zhou, J. H. L. Hansen, and J. F. Kaser, Nonlnear feature based classfcaton of speech under stress, IEEE Trans. Speech, and Audo Processng, Vol. 9, No. 3, 2001, pp [16] Smon Doclo and Marc Moonen, On the Output SNR of the Speech-Dstorton Weghted Multchannel Wener Flter, IEEE SIGNAL PROCESSING LETTERS, Vol. 12, No. 12, [17] Wener, N., Extrapolaton, Interpolaton and Smoothng of Statonary Tme Seres wth Engneerng Applcatons, Wely, Newyork, [18] Wener, N., Paley, R. E. A. C., Fourer Transforms n the Complex Domans, Amercan Mathematcal Socety, Provdence, RI, [19] Koj Ktayama, Masataka Goto, Katunobu Itou and Tetsunor Kobayash, Speech Starter: Nose-Robust Endpont Detecton by Usng Flled Pauses, Eurospeech 2003, Geneva, pp [20] S. E. Bou-Ghazale and K. Assaleh, A robust endpont detecton of speech for nosy envronments wth applcaton to automatc speech recognton, n Proc. ICASSP2002, 2002, Vol. 4, pp [21] A. Martn, D. Charlet, and L. Mauuary, Robust speech / non-speech detecton usng LDA appled to MFCC, n Proc. ICASSP2001, 2001, Vol. 1, pp

6 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 1, [22] Rchard. O. Duda, Peter E. Hart, Davd G. Strok, Pattern Classfcaton, A Wley-nterscence publcaton, John Wley & Sons, Inc, Second Edton, [23] Sarma, V., Venugopal, D., Studes on pattern recognton approach to voced-unvoced-slence classfcaton, Acoustcs, Speech, and Sgnal Processng, IEEE Internatonal Conference on ICASSP'78, 1978, Vol. 3, pp [24] Q L. Jnsong Zheng, Augustne Tsa, Qru Zhou, Robust Endpont Detecton and Energy Normalzaton for Real- Tme Speech and Speaker Recognton, IEEE Transacton on speech and Audon Processng, Vol. 10, No. 3, [25] Harrngton, J., and Cassdy, S., Technques n Speech Acoustcs, Kluwer Academc Publshers, Dordrecht, [26] Makhoul, J., Lnear predcton: a tutoral revew, n Proceedngs of the IEEE 64, 4 (1975), pp [27] Pcone, J., Sgnal modelng technques n speech recognton, n Proceedngs of the IEEE 81, 9 (1993), pp [28] Clsudo Beccchett and Luco Prna Rcott, Speech Recognton Theory and C++ Implementaton, John Wley & Sons. Ltd., 1999, pp [29] L.P. Cordella, P. Fogga, C. Sansone, M. Vento., "A Real- Tme Text-Independent Speaker Identfcaton System", n Proceedngs of 12th Internatonal Conference on Image Analyss and Processng, 2003, IEEE Computer Socety Press, Mantova, Italy, pp [30] J. R. Deller, J. G. Proaks, and J. H. L. Hansen, Dscrete- Tme Processng of Speech Sgnals, Macmllan, [31] F. Owens., Sgnal Processng Of Speech, Macmllan New electroncs. Macmllan, [32] F. Harrs, On the use of wndows for harmonc analyss wth the dscrete fourer transform, n Proceedngs of the IEEE 66, 1978, Vol.1, pp [33] J. Proaks and D. Manolaks, Dgtal Sgnal Processng, Prncples, Algorthms and Aplcatons, Second edton, Macmllan Publshng Company, New York, [34] D. Kewley-Port and Y. Zheng, Audtory models of formant frequency dscrmnaton for solated vowels, Journal of the Acostcal Socety of Amerca, 103(3), 1998, pp [35] D. O Shaughnessy, Speech Communcaton - Human and Machne, Addson Wesley, [36] E. Zwcker., Subdvson of the audble frequency band nto crtcal bands (frequenzgruppen), Journal of the Acoustcal Socety of Amerca, 33, 1961, pp [37] S. Davs and P. Mermelsten, Comparson of parametrc representatons for monosyllabc word recognton n contnuously spoken sentences, IEEE Transactons on Acoustcs Speech and Sgnal Processng, 28, 1980, pp [38] S. Furu., Speaker ndependent solated word recognton usng dynamc features of the speech spectrum, IEEE Transactons on Acoustcs, Speech and Sgnal Processng, 34, 1986, pp [39] S. Furu, Speaker-Dependent-Feature Extracton, Recognton and Processng Technques, Speech Communcaton, Vol. 10, 1991, pp [40] Sddque and M. & Tokh, M., Tranng Neural Networks: Back Propagaton vs. Genetc Algorthms, n Proceedngs of Internatonal Jont Conference on Neural Networks, Washngton D.C.USA, 2001, pp [41] Whteley, D., Applyng Genetc Algorthms to Neural Networks Learnng, n Proceedngs of Conference of the Socety of Artfcal Intellgence and Smulaton of Behavor, England, Ptman Publshng, Sussex, 1989, pp [42] Whteley, D., Starkweather and T. & Bogart, C., Genetc Algorthms and Neural Networks: Optmzng Connecton and Connectvty, Parallel Computng, Vol. 14, 1990, pp [43] Kresmr Delac, Mslav Grgc and Maran Stewart Bartlett, Recent Advances n Face Recognton, I-Tech Educaton and Publshng KG, Venna, Austra, 2008, pp [44] Rajesskaran S. and Vjayalakshm Pa, G.A., Neural Networks, Fuzzy Logc, and Genetc Algorthms- Synthess and Applcatons, Prentce-Hall of Inda Prvate Lmted, New Delh, Inda, [45] N. A. Fox, B. A. O'Mullane and R. B. Relly, The Realstc Mult-modal VALID database and Vsual Speaker Identfcaton Comparson Experments, n Proc. of the 5th Internatonal Conference on Audo- and Vdeo- Based Bometrc Person Authentcaton (AVBPA-2005), New York, Md. Rabul Islam was born n Rajshah, Bangladesh, on December 26, He receved hs B.Sc. degree n Computer Scence & Engneerng and M.Sc. degrees n Electrcal & Electronc Engneerng n 2004, 2008, respectvely from the Rajshah Unversty of Engneerng & Technology, Bangladesh. From 2005 to 2008, he was a Lecturer n the Department of Computer Scence & Engneerng at Rajshah Unversty of Engneerng & Technology. Snce 2008, he has been an Assstant Professor n the Computer Scence & Engneerng Department, Unversty of Rajshah Unversty of Engneerng & Technology, Bangladesh. Hs research nterests nclude bo-nformatcs, human-computer nteracton, speaker dentfcaton and authentcaton under the neutral and nosy envronments. Md. Fayzur Rahman was born n 1960 n Thakurgaon, Bangladesh. He receved the B. Sc. Engneerng degree n Electrcal & Electronc Engneerng from Rajshah Engneerng College, Bangladesh n 1984 and M. Tech degree n Industral Electroncs from S. J. College of Engneerng, Mysore, Inda n He receved the Ph. D. degree n energy and envronment electromagnetc from Yeungnam Unversty, South Korea, n Followng hs graduaton he joned agan n hs prevous job n BIT Rajshah. He s a Professor n Electrcal & Electronc Engneerng n Rajshah Unversty of Engneerng & Technology (RUET). He s currently engaged n educaton n the area of Electroncs & Machne Control and Dgtal sgnal processng. He s a member of the Insttuton of Engneer s (IEB), Bangladesh, Korean Insttute of Illumnatng and Installaton Engneers (KIIEE), and Korean Insttute of Electrcal Engneers (KIEE), Korea.