ROBUST VOICE ACTIVITY DETECTION BASED ON PITCH AND SUB-BAND ENERGY

Transcription

1 ROBUST VOICE ACTIVITY DETECTIO BASED O PITCH AD SUB-BAD EERGY Ziao Zang and Jinlong Lin Scool of Sofware and Microelecronics, Peking Universiy, Beijing 0087, Cina Ziao.Zang@pku.edu.cn, linjl@ss.pku.edu.cn Keywords: Absrac: Voice aciviy deecion, Pic, Sub-band energy crieria. A new Voice Aciviy Deecion (VAD) meod is proposed o rack e various background noises and i can be robus in bo saionary and variable noise environmens. Many previous VAD meods assume a e background only conains cerain kinds of noises, so ey could no deal wi e noise in pracical applicaions efficienly. In proposed approac, deerminae speec, deerminae noise and poenial speec regions are defined. Te firs wo regions are locaed wi exraced pic conour informaion and e ambiguous region will be furer rerieved using updaed resolds of sub-bands energy in obained deerminae noise s frequency domain. Experimens are carried ou wi an exausive comparison o ree sandard VAD meods: G79b, ETSI AFE and AMR. Te resul sows a our approac as a more robus performance an oers in e real circumsances. ITRODUCTIO Voice Aciviy Deecion (VAD) is defined as a procedure o separae speec from silence, noise and oer non-voice segmens. Since i can no only faciliae e speec processing bu also increase e performance of mos recogniion applicaions, VAD as become an essenial fron-end processing sep for various speec signal processing sysems, suc as speec recogniion (L. Karray & A. Marin, 003), speec coding (ITU-T, 997) and speec communicaion (Syed W.Q. & H. Wu, 007). Te imporan saus of VAD sysem for speec signal processing aracs more and more researcers o pay aenion on i. In e early developed VAD algorims, zero crossing raes (ITU-T, 997), linear predicive coding coefficiens, energy resolds (K. Woo e al., 000) and saisical model ave been used. Te Sandard VAD meods G79b (ITU-T, 997) proposed by Inernaional Telecommunicaion Union ITU, Advanced Fron-End: AFE (ETSI, 007) and Adapive Muli-Rae: AMR (3GPP, 00) inroduced by European Telecommunicaion Sandards Insiue ETSI are used on speec coding and communicaion. Tey can acieve a ig speec deecion i rae, bu e non-speec deecion does no efficienly perform as well, especially in e variable background noises. Recenly, e meod (Syed W.Q. & H. Wu, 007) based on an adapive resold relaed o e Signal oise Raio (SR) is proposed. Tis meod can well rack e variable wie noise, babble noise and veicular noise respecively, bu i as a limied es on e voice deecion wi pracical daa. A armonic plus noise model VAD as been inroduced in (E. Fiser e al., 006), wic presened a new pic racking algorim. However, e complex compuaion of is meod makes i ard o use in e real-ime applicaions. Based on e available meods, we find a i is ard o find a robus VAD meod wic can acieve real-ime performance on bo speec and non-speec deecion under e complex siuaion. Tis paper presens a new robus real-ime VAD meod a can rack e noise in real world efficienly. Harmonics, pic value and sub-band energy crieria are inroduced o locae e speec region and rack ime-varying noise respecively wiou raining ime. Firsly, mos vowel segmens wic named deerminae speec region can be deeced by pic measuremen. Ten, sub-band division and energy resolds from deerminae noise region are updaed o rerieve e lef voiced pars in e poenial voice region. Tis paper is organized as follows: Secion presens e principles and framework of proposed meod. In Secion 3, pic measure enancemen 44

2 ROBUST VOICE ACTIVITY DETECTIO BASED O PITCH AD SUB-BAD EERGY are described o find e deermined speec par. Te deailed descripions o explore e lef voiced par from poenial speec region using sub-band energy meod are given in Secion 4. Secion 5 illusraes compared experimenal resuls wi ree sandard VAD meods. Finally, conclusion is given in e las Secion. PROPOSED FRAMEWORK. Principles of Proposed Procedure Te basic principles of proposed algorim wic is referenced from (X.J. Yang & H.S. Ci, 995) are employed in our meod as following: Firs, vowel poneme is an absoluely necessary par in every word for almos all languages. Second, i as been proved a vowel as e periodic naure, wile mos noise and consonan lack of a periodiciy due o eir ineren caracerisics. Terefore, if no pic can be deeced during a cerain long ime, a deerminae noise region mus be conained in is period. Tird, e appearances of speec will increase e energy of signal in eier low frequency or ig frequency. ree segmens C, E and G are defined as e poenial speeces region in wic e consonans ofen exis. Four sub-bands in specrum of deerminae noise F are pariioned o updae e resolds of background noise o deec e exisence of consonans in poenial speec segmens. Comparing e sub-band energies of inervals C, E and G wi ose of deerminae noise F, obvious energy incremen is only deeced in e ail of inerval G beween e wie verical lines in Figure (c). Te segmenaion of is par before inerval H is finised and en is procedure will be repeaed unil e end of is signal. Since e deerminae noise region can be found in nearly every pause during e speec signal, resolds of background noise can be updaed in ime. Terefore, proposed meod can rack e nonsaionary noise efficienly and can be robus for e signal wi complex noise. 3 PITCH MEASUREMET Pic, also called fundamenal frequency, is an effecive feaure o reflec e periodic naure of mos vowels in speec. In is paper subarmonicsummaion (SHS) algorim in (D. Hermes, 988) as been referred due o is fas compuing abiliy, and SHS meod ardly miss all e pic frames in a long speec period. Some improvemens ave been pu forward in proposed meod o increase e accuracy on SHS. 3. SHS Process Figure : Example of proposed procedure.. Framework of Proposed Procedure Based on e principles above, e deailed framework of our approaces is described below: Te signal waveform is sowed in Figure (a). Firsly, pic values are exraced frame by frame unil e firs inerval A wiou pic value wic is longer an 0.75 second comes for as sown Figure (b). Ten e medial region of inerval A is regarded as deerminae background noise marked as F. Te pic s conours can be correcly used o deec e vowels wic is also called deerminae speec region marked as D in Figure (b). Te lef Figure : Pic Measuremen procedure of SHS. Procedure of pic calculaion wi SHS is sown in Figure. Te down-sampled ampliude specrum P(f) in Figure (b) from waveform in Figure (a) is sown in linear frequency abscissa because frequencies above 50 Hz are assumed no o be necessary for e pic-deerminaion process. Ten peak enancemen is applied only o consider e values a and around e peaks. In Figure (c), a 45

3 SIGMAP Inernaional Conference on Signal Processing and Mulimedia Applicaions ransiion from a linear o a logarimic frequency abscissa s is accompanied. Figure (d) only represens sub armonics up o e rank 3. Te sub armonic summaion is calculaed on e sifed logarimic scale log n. Te sub armonic sum specrum H(s) is sowed in formula below. H ( s) = ( n) P( s + log n) () n= Were n is e compression facor, is se o 5, and (n) = 0.84 n-. Te esimaion of pic f= s is e value a wic maximum of e subarmonic sum specrum is calculaed. 3. False Pic Deecion SHS s advanage of low speec missing makes sure e long non-pic par conains a deerminae noise. However, e measuremen may also conain some false pic. For a non-silence signal, ere is always a maximum in e sub armonic sum specrum H(s), even wen signal only consiss of slig noise. Terefore, no all e exraced pices using above meod are vowels frames and ese obained pic value by misake are called false pices. Correlaion coefficien in ime domain as been applied o wipe off e false pices because speec signal ofen as iger correlaion coefficiens an ose of noise. Correlaion coefficien R beween wo signals x i and y i can be calculaed as follows. R = [( xi E ( x) ) ( y i E ( y) )] ( xi E ( x) ) ( y i E ( y) ) i = () E ( x) = x i E ( y ) = y i (3) i = Were x i sars a T seconds before e middle of e signal frame for wic pic was esimaed and y i sars a e middle of e signal. is e sampling number in T=/f seconds and f is e pic value of is frame. Afer all correlaion coefficiens ave been calculaed, a 5- poin median filer is applied o smoo e sequence of coefficiens for e dispersed poins. Te resold for e speec/non-speec decision is se o 0.5. Te exraced pic values wi correlaion coefficiens lower an is resold are regarded as non-pic frames. 4 SUB-BAD EERGY CRITERIA Because ere are no pic value for mos consonans and fracional vowel ails (D.J. Liu & C.T. Lin, 00), sub-band energy crieria for e ambiguous region is adoped o rerieve ese voiced segmens. For vowels, e disribuions of specral energy mainly concenrae in e range of low frequency wile e disribuions for consonans are ofen a ig frequency. Moreover, since mos of e background noises are addiive noises, e arrival of a speec will increase e specral energy of deerminae background noise (A. Davis e al., 006). Terefore, sub-band specral energy crieria in e range of low and ig frequency are inroduced o rerieve e res voice segmens in poenial speec region. 4. Sub-bands Division Te wole frequency band is firsly divided ino wo sub-bands: low-frequency and ig-frequency. Ten eac sub-band is furer divided ino wo sub-bands based on energy disribuion and four sub-bands are pariioned finally. For eac pre-empasized frame in 6 khz sampling rae daa, a 5-poin FFT is applied o obain e ampliude specrum. Tere are 56 specral bins B [0, 55] in 0-8 khz. Firs, e dividing frequency for low and ig frequency is se o 3 khz and e corresponding specral bin is 95. For low frequency range division, assuming specral bin k is e firs division poin, E l and E are e average energy of band B [, k] and B [k, 95]. We calculae eir variances D l and D in low-frequency band as follows: k = ( ( i) ) l k E l 94 D ε (4) D = ( ε ( i) ) 95 E (5) k k ( ) = ε i E p ( i), i =,...,95 (6) p= 0 Were ε (i) is e average energy in low frequency sub-band and E p (i) is e p frame s energy a specral bin i. Te bes division specral bin k on energy disribuion will be acieved wen summaion of D l and D k reaces o minimum since i can divide band 46

4 ROBUST VOICE ACTIVITY DETECTIO BASED O PITCH AD SUB-BAD EERGY according o energy disribuion level. Similarly we can divide e low frequency band ino wo subbands and en e wole frequency becomes four sub-bands: wo sub-bands in low frequency and wo in ig frequency. 4. Sub-band Tresolds Te resold of eac sub-band is se by bo noise energy level and e flucuaion in formula 7: Tr = E max{ d (0)... d ( )}/α (7) d + ( n) = E ( n) E (n=0, =0..3) (8) Were is e frame number of deerminae noise and e resold of sub-band is Tr and α is sensiiviy coefficien beween 0-. d (n) represens e difference beween energy of frame n E (n) and average energy in sub-band E. If e smooed frame s energy can pass e resold, i will be regarded as speec frame. Tis resold reflecs e noise energy level E and is flucuaion max{d (0) d ()}. 5 EXPERIMET 5. Experimen Daa Te experimenal speec signals are colleced from daily conferences recording and Cina Cenral Television (CCTV) programs. Te conference daa were recorded wi PCSA recorder produc and TV programs were recorded wi sofware Pinnacle TVCener Pro including news, enerainmen programs, adverisemens and spors commenaries were golf, ennis and baskeball spors are involved. Te languages spoken in ese daa include Cinese, Englis and Japanese. Te daa are recorded in PCM forma, 8- bi, 6 khz sampling rae wi e oal leng of 35 minues. 5. Experimen Resuls Te es daa are divided ino ree groups: Meeing, TV program and Spors daa. VAD resuls of e proposed meod and ree sandard VAD algorims G79b, ETSI AFE, AMR are given in able. HR means e i rae for bo speec and non-speec deecion a frame level. HR0 and HR are defined as speec i-rae and non-speec Table : Compare proposed meod o sandard VAD. Daa Leng Resul G79b AFE AMR Proposed Meeing 65min TV program 5min Spors 35min HR 9.% 8.4% 86.8% 97.% HR0 98.9% 99.3% 99.7% 99.% HR 84.9% 49.% 58.9% 93.0% HR 84.8% 83.7% 84.6% 94.8% HR0 93.4% 99.% 99.8% 98.7% HR 67.9% 5.3% 53.7% 87.% HR 8.6% 80.4% 79.4% 9.4% HR0 88.9% 96.3% 98.7% 95.5% HR 69.3% 47.8% 40.% 8.8% deecion i-rae a frame level respecively. From able, we can see a speec deecion i rae (HR0) of our proposed meod for ese es daa can reac 99.%, 98.7% and 95.5% respecively, wic is iger an a of G79b and a lile bi lower an ose of AFE and AMR. Moreover, proposed meod reaces o e mos efficien nonspeec deecion i rae (HR). Taking Meeing daa as example, HR of our meod can reac 93.0% wile 84.9%, 49.% and 58.9% for G79b, AFE and AMR respecively. Resul sows a our meod acieves bes HR for overall speec and non-speec deecion. To beer explore e performance of our meod, deailed comparison of HR, HR0 and HR in ese four meods under difference SR is given in Figure 3. Figure 3: Resuls comparison under differen SR. I can be seen from Figure 4(a) a ese four meods can acieve more an 97% speec i rae (HR0) wi SR values beween 5 and 0dB. Wi e increasing of noise level, G79b s HR0 decreases below 85% wile our meod and AFE, AMR is sill above 90%. In Figure 4(b), our meod obains 5%~0% improvemens an G79b and 0%~50% improvemens an AFE and AMR. In Figure 4(c), experimenal resuls sow a proposed meod improves more an 0% overall i rae of 47

5 SIGMAP Inernaional Conference on Signal Processing and Mulimedia Applicaions speec and non-speec an ese ree sandard VAD meods even wen ese recordings are dominaed wi speec over non-speec. Te experimens are esed on a Core.3GHz Windows XP PC and e real-ime rae of proposed meod is /50. I proves a proposed meod is compeen for real-ime speec processing sysem. 6 COCLUSIOS Tis paper presens a robus VAD meod based on pic measuremen and sub-band energy. Te proposed meod ouperforms G79b, AMR and AFE under differen SR environmens, especially on e non-speec deecion. Experimenal resul sows proposed meod is more reliable in e pracical circumsances. K. Woo, T. Yang, K. Park, and C. Lee, 000. Robus voice aciviy deecion algorim for esimaing noise specrum, Elecronics Leers, pp L. Karray and A. Marin, 003. Toward improving speec deecion robusness for speec recogniion in adverse environmens, Speec Communicaion, pp Syed W.Q., Hsiao-Cun Wu, 007. Speec waveform compression using robus adapive voice aciviy deecion for nonsaionary noise in mulimedia communicaions, Global Telecommunicaions Conference, pp X.J Yang., H.S. Ci, 995. Speec signal digial processing, Elecronic Indusry Press, Beijing. ACKOWLEDGEMETS My friend Kun Liu gives grea elp during e wole researc and e Figure in is paper is modified by e picure from D. Hermes aricle, auor gives a grea appreciaion o bo of em. REFERECES 3GPP, 00. Speec codec speec processing funcions; Adapive Muli-Rae-Wideband (AMR-WB) speec codec; Voice Aciviy Deecor (VAD). A. Davis, S. ordolm, S.-Y. Low, R. Togneri, 006. A muli-decision sub-band voice aciviy deecor, Proceedings of EUSIPCO, Florence Ialy. D. Hermes, 988. Measuremen of pic by subarmonic summaion, Te Journal of e Acousic Sociey of America, pp Der-Jenq Liu, Cin-Teng Lin, 00. Fundamenal frequency esimaion based on e join imefrequency analysis of armonic specral srucure, IEEE Trans. Speec Audio Process, pp E. Fiser, J. Tabrikian, S. Dubnov, 006. Generalized likeliood raio es for voiced-unvoiced decision in noisy speec using e armonic model, IEEE Transacions on Audio, Speec and Language Processing, pp ETSI, 007. Speec processing, ransmission and qualiy aspecs (STQ); Disribued speec recogniion; Advanced fron-end feaure exracion algorim; Compression algorims, ETSI ES Recommendaion. ITU-T, 997. A silence compression sceme for G.79 opimized for erminals conforming o recommendaion V.70, ITU-T Rec. G. 79, Annex B. 48