Speaker Verification Reinforced by Objective Wavelet Packets-based Speech Parameterization

Transcription

1 Speaker Verificatio Reiforced by Obective avelet Packets-based Speech Parameterizatio Mihalis Siafarikas, Todor Gachev, ad Nikos Fakotakis ire Commuicatios Laboratory, Uiversity of Patras, Rio-Patras 26500, Greece Abstract. I attempt to ehace the discrimiatio ability of preset speaker verificatio systems, we study alterative ways for parameterizatio of huma voice. Utilizig the flexibility provided by wavelet packet aalysis, ad rederig a accout of the most recet psycho-acoustical studies, the authors ivestigate i a obective way the relative importace of costituet disoit frequecy subbads of speech sigal. I extesive experimetatio, the cotributio of each subbad i relatio to its correspodig frequecy resolutio is quatified. Takig a advatage of the isight gaied i these experimets, we costruct a ovel wavelet packet-based speech feature set that reders a accout of the relative importace of the frequecy subbads. I comparative experimets performed o 2001 NIST Speaker Recogitio Evaluatio database, cotrastive evaluatio of the proposed speech features, Mel-frequecy cepstral coefficiets (MFCCs), as well as previous wavelet packet based features that were reported successful for speaker recogitio is performed. 1 Itroductio Although cepstrum-based speech features, such as MFCCs, are curretly the most popular choice for speech recogitio, it is doubtful that they are the best choice for speaker recogitio give the dissimilar requiremets for these two tasks. Clearly, speech recogitio aims at capturig typical properties of phoemes, irrespective of ay speaker-origiated variatios i the speakig style, while speaker recogitio aims at exploitig these idividualities. These distictive traits of the speaker recogitio tasks motivated the authors of the preset study to search for a more geeral approach (tha the Mel-scale related speech features) that better demarcates the idividual voices. As a alterative to the traditioal Fourier Trasform based techiques for aalyzig time series, avelet Packet Trasform (PT) has bee prove a effective sigal-processig tool i a variety of speech processig applicatios. Specifically, PT has bee used as a close approximatio of the Mel-frequecy scale for the purpose of speech recogitio i [1] ad [2], i lieu of Discrete Fourier Trasform (DFT). Furthermore, i [], Sarikaya et al. used the wavelet features itroduced i [1] for speaker idetificatio referrig that they outperformed the MFCCs. I [4], the authors, relyig o DFT aalysis, explore the discrimiatio ability of differet fre- GESTS Iteratioal Trasactio o Speech Sciece ad Egieerig, SuJi Publishig Co., Volume1, Number 2, December pp URL:

2 quecy subbads, istead of strictly adherig to the Mel-scale. Results from subective speaker verificatio tests that study a umber of frequecy subbads, were used for the derivatio of perceptually motivated features, which reportedly outperformed the MFCCs. I the preset work, employig wavelet packet aalysis, the authors study i a obective maer the relative importace of eight disoit frequecy subbads of speech sigal by comparative evaluatio of the speaker verificatio performace. The results of that study are used to costruct a wavelet packet tree that effectively represets the huma voice idividuality. Based o the proposed wavelet packet tree, a ovel speech features set, amed Obective avelet Packet Features (OPFs) that is fie-tued for the task of speaker verificatio was created. Eve though our work was ispired by [1], [2], [], ad [4], there are maor differeces betwee these earlier studies ad our approach. Apart from usig differet wavelet ad correspodig cougate mirror filters, the authors of [1], [2], ad [] have used wavelet packets for approximatig the Mel-frequecy scale, while we utilize the wavelet packets for achievig a more effective divisio of the frequecy scale that exploits the relative importace of the frequecy subbads. The pricipal differece betwee [4] ad our work is that we do ot deped o subective speaker verificatio tests but o the cotrary, we perform a obective study of the relative importace of the disoit frequecy subbads. Besides the aforemetioed differeces, we do ot cofie to frequecy aalysis usig DFT that provides a fixed tilig of the time-frequecy plae, ad thus, requires further weighig of the frequecy subbads to facilitate a perceptually based frequecy scale. Istead, we exploited the flexibility provided by wavelet packets as cocers the tilig of the time-frequecy plae i order to follow the discrimiatio performace of differet subbads, accordig to results of our obective study. These results guided us to attribute differet resolutios to separate frequecy subbads buildig accordigly a wavelet packet tree that provided the proposed features set, OPFs. The experimetal results preseted i Sectio 5 provide compariso betwee the proposed OPFs ad MFCCs, as well as with the wavelet packet features itroduced i [1], [2], ad []. 2 avelet Packet Aalysis avelets are fuctios with specific properties useful i the aalysis of cotiuous ad discrete-time sigals. This sectio presets wavelet aalysis payig particular attetio to its trasforms o discrete data, Discrete avelet Trasform (DT) ad Discrete avelet Packet Trasform (DPT). 2.1 Discrete avelet Trasform The basic compoet of wavelet aalysis is Cotiuous avelet Trasform (CT). It provides a time-scale represetatio of a cotiuous fuctio usig traslatios ad dilatios of a sigle fuctio, called wavelet. A aalytic presetatio of wavelet 61

3 aalysis ad its basic trasform, CT, ca be foud i [9]. The basic tool for the practical aalysis of discrete-time sigals via wavelets is the Discrete avelet Trasform (DT). It is a orthoormal trasform that ca be regarded as a sub-samplig of the two dimesioal CT o dyadic scales 2, s = N, ad times t = k 2, k Z, providig a oe dimesioal time-scale represetatio of the sigal. However, DT ca be formulated etirely i its ow right. Thus, the DT of level J N of a discrete-time sigal x, [ ] = 0,, N 1 with N = 2 J, is a N dimesioal vector, = V 1 2 J J. Each is a dimesioal vector of wavelet coefficiets each oe of which is associated with adacet N /2 time itervals of width 2 ad frequecy iterval + 1/2 1,1/2 while V J is a oe dimesioal vector cotaiig the scalig coefficiet associated with the whole time J 1 iterval of width 2 ad the frequecy iterval 0,1/ 2 J +. Therefore, the DT provides a octave based decompositio of the frequecy domai of a sigal i the way depicted i the Fig. 1. The rectagles therei idicate the optimal resolutio achieved by the time-frequecy capability of DT. I practice, DT is computed by the successive applicatio of two discrete filters, called wavelet filter ad scalig filter, iitially o the sigal, ad subsequetly o the lower frequecy part of the resultig sigal, followed by sub-samplig by a factor of two. A filter { h : l = 0, L 1, L = 2 k, k N is called a wavelet filter if L 1 l= 0 h = 0 ad l l L 1 l = 0 hh = δ l l+ 2,0 T. As its ame suggests, there exists a explicit coectio betwee this filter ad the wavelet fuctio used i CT [15]. The scalig filter is defied i terms of the wavelet filter via the quadrature mirror relatioship ( 1) l + g 1 l = hl 1 l. Applicatio of the wavelet filter { h l } is equivalet to selectig the higher frequecy part of a sigal, while applicatio of the filter { g l } is equivalet to selectig the lower frequecy part. Therefore, i the decompositio of a sigal with the DT, oly the lower frequecy bad is decomposed, givig a right recursive biary tree structure, where its right child represets the lower frequecy bad ad its left child represets the higher frequecy bad. Ideally, it would be desirable to have wavelet ad scalig filters the frequecy respose of which were limited to the frequecy bads [ 0,1/ 4 ] ad [ 1/4,1/2 ], respectively. I this case, the actual time-frequecy tilig of the DT would coicide with the optimal tilig depicted i Fig. 1. The degree that these omial frequecy itervals are approximated (by the actual frequecy decompositio achieved by DT) depeds o the frequecy characteristics of the wavelet ad scalig filters. Therefore, although the optimal time-frequecy aalysis of DT is depicted i Fig. 1, differet wavelet filters ca provide fie-tued DT aalyses. Furthermore, wavelet ad scal- } 62

4 ormalized frequecy 1/2 N/2 time itervals 1 1/4 1/8 1/16. 1/2 J 1/2 J+1. N/4 time itervals N/8 time itervals. N/2. N 2 J V J discrete time Fig. 1. DT time frequecy aalysis ig filters play a similar role i the formulatio of Discrete avelet Packet Trasform (DPT) preseted i Sectio Discrete avelet Packet Trasform Discrete avelet Packet Trasform (DPT) is a geeralizatio of the DT that allows a effective represetatio of the time frequecy properties of a discrete sigal so that useful features for a particular purpose ca be appropriately extracted. I this subsectio, we preset the aalysis of a time series with DPT showig the advatages over DT as cocers their capability for time-frequecy represetatio. J Let x [ ], = 0, N 1, where N is a iteger multiple of 2 for some positive iteger J, deote a real valued discrete time sigal. For 0 J, the level DPT of x [ ] is a orthoormal trasform yieldig a N dimesioal vector of coefficiets that ca be partitioed as, where each is a N /2 dimesioal vector, each elemet of which is omially associated T + 1 with adacet time itervals of width 2 ad frequecy iterval I =, Thus, these 2 vectors divide the Nyquist frequecy iterval [ 0,1/ 2 ] ito 2 itervals of equal width ad each oe of its N /2 elemets provide iformatio associ- ated with the time itervals k 2, ( k 1) 2 +, k = 0, N /2 1. Therefore, DPT provides localized time-frequecy iformatio about a sigal. I additio, each level of DPT provides homogeeous frequecy ad time aalysis, as show i Figure 2, i cotrast to DT that provides a octave-based decompositio. 6

5 ormalized frequecy 1/2 N/2 time itervals 7 N/2 time itervals 6 N/2 time itervals 5 1/4 N/2 time itervals N/2 time itervals 4 1/8 1/16 2 N/2 time itervals N/2 time itervals N/2 time itervals N discrete time Fig. 2. Time frequecy aalysis achieved with level = of DPT The most appealig aspect of DPT is that carefully selected basis vectors belogig to differet level DPTs ca be grouped together i order to create a eve larger collectio of orthoormal trasforms. This is achieved by orgaizig all the DPTs for levels = 0,1, J ito a tree structure, called wavelet packet (P) tree. That is possible sice, i goig from the level 1 DPT to the oe of ext level 1, each vector is circularly filtered ad dow-sampled twice: oce with the wavelet filter { h l } ad oce with the scalig filter { g l } = 2 = 2 +1, yieldig two vectors idexed by ad. Havig costructed the P tree, the coefficiet vectors ca be collected together to form a set { } S = : = 0, J, = 0, 2 1, where each S is omially associ- ated with the frequecy bad. Ay subset S S that provides a o-overlappig complete coverage of I 1 [ 0,1/ 2 ] with coefficiet vectors yields a orthoormal DPT. I this way, DPT provides a flexible tilig of the time frequecy plae with various frequecy resolutios i the correspodig time itervals. The DPT decompositio, as a orthoormal trasform, uses the followig relatioship: x 2 [ ] S1 to partitio the eergy i discrete sigal x tilig provided by the particular DPT. 2 =, [ ] accordig to specific time-frequecy 64

6 Relative importace of frequecy subbads I order to determie the maximum frequecy resolutio ecessary to capture the speaker idetity, we took ito cosideratio the cocept of critical badwidth itroduced by Fletcher i [5]. Zwicker i [6] estimated that the critical badwidth is costat at 100 Hz for cetre frequecies up to 500 Hz, while for higher frequecies the badwidth icreases approximately i proportio with cetre frequecy. However, more recet experimets [7] have provided evidece that the critical badwidth ca be as arrow as 0 Hz for frequecies below 500 Hz. The last led us to costruct a DPT of level = 7 sice the critical badwidth of 0 Hz ca be effectively approximated by a resolutio of 1.25 Hz = 8000 Hz. (1/2). (1/2 7 ), where 8000 Hz is the samplig frequecy. ith the itetio to study the sigificace of each frequecy subbad cocerig the differetiatio of the speakers voices, we performed a umber of speaker verificatio experimets utilizig the Polycost speaker recogitio corpus [10]. I total, speech from all the 74 male speakers were employed, each represeted with his first three sessios, ad with te seteces per sessio. The full Nyquist frequecy rage of [0, 1/2] of the speech sigal is separated ito eight disoit subbads of width (1/2 ). (1/2) = 1/16. These subbads ad their correspodig ceter frequecies are show i the first two colums of Table 1. The badwidth of 1/16 was chose with the aim of obtaiig a smoothed estimatio of the relative importace of the correspodig subbad avoidig the threshold effect whe dealig with smaller sectios. Istead of performig subective speaker verificatio tests, as accomplished i [4], we oppositely evaluated the speaker verificatio performace by employig the automatic speaker verificatio system described i [1]. However, i cotrast to [1], where MFCCs speech features are cosidered, here, speaker-specific iformatio available i each tested frequecy subbad was aalyzed usig DPT of depth = 7, correspodig to a frequecy resolutio of 1.25 Hz, ad selectig the coefficiets represetig this particular frequecy bad, as show i Figure 1. Speech sigal was preprocessed as described i Sectio 4. The relative importace of spectral iformatio i each of the subbads, show i colum 1 of Table 1, was evaluated i the sese of Equal Error Rate (EER) for the above metioed frequecy resolutio of 1.25 Hz, with the results preseted i the third colum of the same table. Table 1. The EER for the 8 disoit frequecy subbads, alog with the proposed resolutios. Frequecy Iterval Ceter Frequecy EER at resolutio 1/2 8 [%] Proposed resolutio [Hz] [0, 1/16] 1/ [1/16, 1/8 ] / [1/8, /16] 5/ [/16, 1/4 ] 7/ [1/4, 5/16] 9/ [5/16, /8 ] 11/ [/8, 7/16] 1/ [7/16, 1/2 ] 15/

7 4kHz 2kHz 1kHz 500Hz 250Hz 125Hz 62.5Hz 1.25Hz HPF LPF 2 2 Fig.. The proposed wavelet-packet tree. Oly the frequecy bads draw with solid bold lie are utilized i the computatio of the speech features As it ca be observed i Table 1, the smallest EER is attaied i the lowest two frequecy bads [0, 1/16] ad [1/16, 1/8], which led us to coclude that most speakerspecific iformatio is revealed i these two frequecy bads. The frequecy bads [1/8, /16], [1/4, 5/16], ad [/8, 7/16] have lower EER, ad thus appear to cotai more speaker-specific iformatio tha the frequecy bads [/16, 1/4], [5/16, /8], ad [7/16, 1/2] that demostrated the worst results. Apart from revealig useful speaker-specific iformatio cocerig the differet subbads, our approach exploits these results i a costructive way to improve the discrimiative resolutio of the wavelet packet aalysis. Havig i mid that the DPT at depths 7, 6, ad 5, result i 16, 8, ad 4 frequecy itervals of width 1/2 8, 1/2 7, ad 1/2 6, respectively, i each 1/16 subbad ad cosiderig the aalysis of the speaker verificatio performace of each frequecy subbad, we built a wavelet packet tree as follows: The frequecy bads were clustered ito groups accordig to the observed EER, i.e., {[1, 1/16], ad [1/16, 1/8]}, {[1/8, /16], [1/4, 5/16], ad [/8, 7/16]} ad {[/16, 1/4], [5/16, /8], ad [7/16, 1/2]}. Due to their differet relative importace, these groups are expected to cotribute uequally ito the fial vector of features. Thus, we propose that the frequecy resolutios should rely o the relative importace as it is show i the last colum of Table 1. By implicatio, i the feature extractio stage preseted i the ext Sectio 4, sixtee coefficiets are take ito accout for each of the frequecy subbads i the first group, eight coefficiets for 66

8 the frequecy subbads i the secod group, ad four coefficiets for the frequecy subbads i the third group. This, i tur, results i a represetatio of the frequecy subbads accordig to the respective EERs: more coefficiets for the lowest EER, less coefficiets for the higher EER, ad eve less coefficiets for the highest EER. Ito the block diagram preseted i Fig. 4, that step is referred to as "Obective Subbad Eergy Estimatio". Due to their lowest EER, ad thus a relatively higher importace, the lowest two frequecy subbads receive the highest frequecy resolutio of 1/2 8. The subbads with EER betwee 21% ad 24% receive a frequecy resolutio of 1/2 7, ad the subbads with EER higher tha 24% receive the lowest resolutio of 1/ Feature Extractio Based o the relative importace of disoit frequecy subbads derived i Sectio, we costructed a wavelet packet tree that effectively icorporates the idividual cotributio of each frequecy subbad i the feature extractio process. (The lowest four bis of width 1/2 8, coverig the frequecy rage [0, 1/64] were excluded from the wavelet tree because that frequecy iterval is very much affected by the characteristics of trasmissio chaels.) Thus, the wavelet packet tree show i Fig. was costructed providig a total of B=64 frequecy subbads. Usig the otatios itroduced i Sectio 2, the specific wavelet tree is described as a subset of DPT vectors, i.e , 6 6, 5 5, S1 = , 5 5, 6 6, 5 5 The proposed speech features, OPFs, are computed as depicted i the block diagram of Fig. 4: The speech sigal is filtered by a fifth order Butterworth filter with pass-bad from 80 Hz to 800 Hz, followed by framig i itervals of 2 millisecods, with a skip rate of 16 millisecods. Due to the compact support of wavelets, o Hammig or other complex widow is required, ad therefore a rectagular oe is 1 cosidered. A pre-emphasis filter Hz () = z is employed. A voiced/ uvoiced frame decisio is obtaied usig a pitch estimatio based o the modified autocorrelatio method with clippig [8]. Oly those feature vectors represetig voiced speech frames are further used to represet the speaker s idetity. Next, DPT of level = 7 is applied to the voiced speech frames i accordace to the wavelet tree referred to i Sectio, providig a total of B=64 frequecy subbads, as depicted i Fig.. I order to avoid appearace of false large amplitude coefficiets at the margis of every speech frame, boudary wavelets were utilized i the computatio of the DPT. The, the eergy i each frequecy bad is computed, ad divided by the total umber of coefficiets preset i that particular bad. More specifically, the subbad sigal eergies are computed for each frame as, 67

9 Speech Sigal SAMPLING 8 KHZ FILTERING HZ PRE- EMPHASIS AVELET PACKET DECOMPOSITION OBJECTIVE SUBBAND ENERGY ESTIMATION VOICED/ UNVOICED DECISION AMPLITUDE LOGARITHMIC COMPRESSION FRAMING DISCRETE COSINE TRANSFORM Feature Vector Fig. 4. Computatio of the proposed wavelet packet features (OPFs) N /2 k 2 [ [] i ] Ep = i= 1 k, S1, p = 1,..., B, N /2 k k where [ i] is the i-th coefficiet of the DPT vector. Fially, a logarithmic compressio is performed ad a Discrete Cosie Trasformatio is applied o the logarithmic subbad eergies i order to obtai decorrelated coefficiets: B ip ( 1 2) Fi ( ) = log 1 10 Ep cos( ), i= 1,..., r, p= B where r is the umber of feature parameters. I order to reduce the dimesioality of the feature vector, we compute oly the first 5 coefficiets, sice they represet 99.99% of the eergy of the complete set of Experimets ad Results The speaker verificatio system [1] was used as a platform to evaluate several sets of speech features. I the comparative experimets, we have utilized the male part of the 2001 NIST SRE corpus [11]. Approximately 40 secods of voiced speech for traiig of each user model were available. The commo referece model was built from oe hour ad forty miutes of voiced speech, exploitig the male traiig data offered i the 2002 NIST SRE database [14]. Each speaker verificatio experimet icluded 850 target ad 8500 impostor trials with legths betwee 0 ad 60 secods, coverig the etire diversity of trasmissio chael types, defied i the complete oe speaker detectio task [11]. I the experimets with the MFCC features, we adhered to a approximatio of the Mel-scale with 2 filters, referred to as MFCC FB-2, which covers the frequecy diapaso Hz. It was foud that MFCC computed from 2 filters are more successful for speaker verificatio applicatios, whe compared to other implemetatios for example: with a filter-bak of 20 filters as implemeted i the HTK [12]. The authors presume the better performace of MFCC FB-2 is due to the higher spectral resolutio of these speech features, whe compared to the traditioal MFCC FB-20. The MFCCs FB-2 were computed as described i [1]. The Farooq-Datta s 68

10 features ad the Sarikaya s features were estimated by followig the methodology of the correspodig author: [2] ad [1], respectively. The proposed wavelet packet based features set, OPFs, has bee computed as described i Sectio 4. Table 2 reveals a compariso amog the wavelet packet based features ad MFCCs. Colum 2 of Table 2 presets the actual coefficiets for each type of speech features, icluded i the correspodig experimet, while colum provides the obtaied EERs. The best set for each kid is deoted by the symbol *. I the case of MFCCs, excludig the first cepstral coefficiet of the feature vectors reduced the ifluece of trasmissio chael mismatch betwee trai ad test, ad thus decreased the EER. It was foud that removig the ext cepstral coefficiets after the first oe deteriorates the speaker verificatio performace. However, for all wavelet packets based features studied here, amely OPFs, P2, ad P, it was observed that excludig the first three coefficiets from the feature vectors leads to sigificat reductio of the EER. Table 2: The EER i percetage for the evaluated feature sets Speech Features Set of Coef. EER [%] OPF proposed {1,2,,5} OPF proposed {2,,,5} OPF proposed {,4,,5} * OPF proposed {4,5,,5} 1.98 P2 Sarikaya {1,2,,24} P2 Sarikaya {2,,,24} P2 Sarikaya {,4,,24} * P2 Sarikay a {4,5,,24} 15.9 MFCC FB-2 {1,2,,2} * MFCC FB- 2 {2,,,2} 16.5 MFCC FB-20 (H TK) {2,,1,, } P Farooq-Datta {1,2,,1} P Farooq-Datta {2,,,1} P Farooq-Datta {,4,,1} * P Farooq-Dat ta {4,5,,1} he we compare the best members of each kid of speech features, show i Table 2, the Farooq-Datta s set, P, exhibits the highest error rate, while the proposed set, OPF, expresses the lowest oe. The Sarikaya s features, P2, were cofirmed to perform better tha the MFCCs, but are outperformed by the proposed features, OPFs. I coclusio, we ca geeralize that the experimetal results cofirmed the advatage of proposed speech features, OPFs. Optimized i a systematic way to emphasize huma voice idividuality, they provide a better speaker verificatio performace tha the MFCCs, ad the Farooq-Datta s ad Sarikaya s wavelet packets based speech feature sets. 69

11 6 Coclusio A ovel, wavelet packet based speech features set, appropriate for speaker verificatio, was proposed. Our cotributio is maily i the wavelet packet tree desig that was costructed accordig to results from a obective study of the relative importace of each frequecy subbad i the demarcatio of huma voices. Thus, the speech feature set we propose is fie-tued to emphasize the relatively more importat spectral subbads for voice differetiatio. A comparative experimetal evaluatio of the proposed features, performed o a well-kow speaker recogitio corpus, proved the practical sigificace of our approach. The proposed speech features demostrated a superior performace, whe cotrasted to other wavelet packet-based features ad to the Mel-frequecy scaled cepstral coefficiets, due to a better represetatio of the speaker-specific variatios of the speech sigals. Refereces 1. Sarikaya, R., Hase, H.L.: High Resolutio Speech Feature Parameterizatio for Moophoe-based Stressed Speech Recogitio, IEEE Sigal Processig Letters, vol.7, o.7 (2000) Farooq, O., Datta, S.: Mel-scaled avelet Filter Based Features for Noisy Uvoiced Phoeme Recogitio, Proc. of ICSLP 2002, Dever, Colorado, USA (2002) Sarikaya, R., Pellom, B.L., Hase, H.L.: avelet Packet Trasform Features with Applicatio to Speaker Idetificatio, Proc. of IEEE Nordic Sigal Processig Symp., Visgo, Demark (1998) Orma, O.D., Arsla, L.M.: Frequecy Aalysis of Speaker Idetificatio, Proc. of The Speaker Recogitio orkshop 2001: A Speaker Odyssey, Crete, Greece, (2001) Fletcher, H.: Auditory Patters, Reviews of Moder Physics, o. 12 (1940) Zwicker, E.: Subdivisio of the Audible Frequecy Rage ito Critical Bads (Frequezgruppe), The J. of Acoustical Society of America, vol. (1961) Moore, B.C.J.: A Itroductio to the Psychology of Hearig, Academic Press, Lodo, 5th ed. (200) 8. Rabier, L.R., Cheg, M.J., Roseberg, A.E., McGoegal, C.A.: A Comparative Performace Study of Several Pitch Detectio Algorithms, IEEE Tras. o ASSP, Vol. ASSP-24, No.5 (1976) Mallat, S.: A wavelet Tour of Sigal Processig, Academic Press, Sa Diego, USA (1998) 10. Petrovska, D., Heebert, J., Meli, H., Geoud, D.: Polycost: A Telephoe-Speech Database for Speaker Recogitio, Speech Commuicatio, vol.1, o.2- (2000) The NIST Year 2001 Speaker Recogitio Evaluatio Pla, The NIST of USA (2001) Available: Youg, S.J.: The HTK Hidde Markov Model Toolkit: Desig ad Philosophy, Techical Report TR.15, Departmet of Egieerig, Cambridge Uiversity, UK (199) 1. Gachev, T., Fakotakis, N., Kokkiakis, G.: Text-Idepedet Speaker verificatio Based o Probabilistic Neural Networks, Proc. of Acoustics, Patras, Greece (2002) The NIST Year 2002 Speaker Recogitio Evaluatio Pla, The NIST of USA (2002) Available: Percival, D.B., alde, A.T.: avelet Methods for Time Series Aalysis, Cambridge Uiversity Press, USA (2000) 70

12 Biography ame: Mihalis Siafarikas Address: ire Commuicatios Laboratory, Uiversity of Patras, Rio-Patras, Greece. Educatio & ork experiece: Dipl. Egieer degree i Telecommuicatios & Electroics Egieerig from Helleic Air Force Academy, i MSc i Satellite Commuicatios Egieerig from Uiversity of Surrey, UK, i Mathematics degree & MSc i Applied Mathematics from Uiversity of Patras, Greece, i 2002 ad 2004, respectively. I March 200, he started his Ph.D. study with the ire Commuicatios Laboratory i the Speaker & Laguage Recogitio area. Tel: msiaf@wcl.ee.upatras.gr ame: Todor Gachev Address: ire Commuicatios Laboratory, Uiversity of Patras, Rio-Patras, Greece. Educatio & ork experiece: Dipl. Egieer degree i Electrical Egieerig from the Techical Uiversity of Vara, Bulgaria, i 199. From February 1994 to August 2000, he cosequetly occupied egieerig, research, ad teachig staff positios at the same uiversity. Sice September 2000, he is with the ire Commuicatios Laboratory, where i February 2001 he started his Ph.D. study i the Speaker Recogitio area. Tel: tgachev@wcl.ee.upatras.gr ame: Nikos Fakotakis Address: ire Commuicatios Laboratory, Uiversity of Patras, Rio-Patras, Greece Educatio & ork experiece: B.Sc. degree i Electroics from the Uiversity of Lodo (UK) i 1978, M.Sc. degree i Electroics from the Uiversity of ales (UK), ad Ph.D. degree i Speech Processig from the Uiversity of Patras, Greece, i Sice 1986, cosequetly he occupied Lecturer, Assistat Professor, ad Associate Professor positios. Curretly he is a Professor i the area of Speech ad Natural Laguage Processig ad Head of the Speech ad Laguage Processig Group at the ire Commuicatios Laboratory. Tel: fakotaki@wcl.ee.upatras.gr 71