Improved Codebook-based Speech Enhancement based on MBE Model

Transcription

1 INTERSPEECH 7 August 4, 7, Stochol, Sweden Iproved Codeboo-based Speech Enhanceent based on MBE Model Qizheng Huang, Changchun Bao, Xianun Wang Speech Audio Signal Processing Laborator, Facult of Inforation Technolog, Being Universit of Technolog, Being, China, 4 huangqizheng@eails.bjut.edu.cn, baochch@bjut.edu.cn, b4@eails.bjut.edu.cn Abstract This paper provides an iproved codeboo-based speech enhanceent eod using ulti-b ecitation (MBE) odel. It ais to reove e noise between e haronics, which a eist in codeboo-based enhanced speech. In general, e proposed sste is based on analsis-wisnesis (AwS) fraewor. During e analsis stage, acoustic features are etracted including pitch, haronic agnitude voicing fro nois speech. These paraeters are obtained on e basis of e spectral agnitudes obtained b codeboo-based eod. During e snesis stage, different snesis strategies for voiced unvoiced speech are eploed. Besides, is paper introduces speech presence probabilit to odif e codeboo-based Wiener filter so at ore accurate acoustic paraeters can be obtained. The proposed sste can eliinate noise not onl between e haronics, but also in e silent segents, especiall in low SNR noise environent. Eperients show at, e perforance of e proposed eod is better an traditional codeboo-based eod for different tpes of noise. Inde Ters: speech enhanceent, MBE odel, codeboos, analsis-wi-snesis, acoustic features. Introduction Speech enhanceent plas an iportant role in speech counication due to its wide applications. But it still faces enorous challenges because of e variabilit of noise. Conventional speech enhanceent eods such as spectral subtraction [], Wiener filter [] statistical odel-based eods [3] perfor well for e stationar noises, but e perforance of e degrades rapidl in e non-stationar noise environent. Codeboo-based eod stores e prior inforation of speech noise spectral envelope in e codeboos. The concerned short-ter predictor (STP) paraeters including AR coefficient AR gain are estiated online. In [4, 5], STP paraeters are first estiated b aiu-lielihood Baesian MMSE eod, respectivel. Then Wiener filter is constructed using e estiated paraeters. The codeboos contain a priori inforation about speech noise, erefore e enhanced speech of e codeboo-based eod has a saller spectral distortion, especiall in non-stationar noise conditions. But e codeboo-based eod is onl used to odel e spectral envelope, so ere is a lot of residual noise between adjacent haronics. Furerore, ere is a large aount of residual noise in e silent segents due to e inaccurac of STP paraeters. In recent ears, e analsis-wi-snesis (AwS) eod is a hotspot of speech enhanceent. It has two stages, nael analsis stage snesis stage. During e analsis stage, acoustic features are etracted fro speech signals, such as pitch haronic agnitude. During e snesis stage, enhanced speech is snesized b e features. Copared wi conventional speech enhanceent algoris, AwS eods can preserve e haronic structure better eliinate ore noises. In [6], a speech enhanceent eod which is based on pure speech features reconstruction is proposed. The sste uses e aiu a posteriori eod to estiate clean spectral envelope fro nois spectral envelope, e enhanced speech is snesized using sinusoidal odel. In addition, haronic plus noise odel [7] also uses e AwS fraewor. It etracts four auditor paraeters including pitch, spectral envelope, spectral gain, voicing ied function. The spectral envelope is obtained b tracing line spectru frequenc trajector rough Kalan filtering. Aong e AwS eods, MBE is a significant odel. It taes into account e different wa of speech production for voiced portion unvoiced portion, so high qualit speech can be snesized. This paper is based on e MBE odel [8] codeboobased eods [4, 5]. The proposed sste uses AwS fraewor ais at reoving noise between e haronics generating clean haronics. At first, e linear prediction (LP) paraeter codeboos of speech noise are trained offline. Then e initial enhanceent spectru is obtained b codeboo-based eod. Subsequentl, e acoustic features, such as pitch, haronic agnitude voiced/unvoiced (V/UV) decision are etracted fro e enhanced spectru. Finall, e voiced unvoiced speech signals are snesized b different strategies. The reainder of e paper is organized as follows. Section eplains e signal odel Baesian codeboobased speech enhanceent eod. Section 3 eplains e proposed fraewor in detail. Eperients results are shown in Section 4 finall, Section 5 concludes e paper.. Codeboo-based Baesian eod We consider an additive noise odel, speech noise are independent, n ( ) n ( ) wn ( ) () n ( ), n ( ) wn ( ) represent e sapled nois speech, clean speech noise, respectivel. As a priori inforation, e spectral shape codeboos of clean speech noise are trained offline in advance, respectivel. The ai of codeboo-based eod is to estiate STP paraeters, nael LP coefficient ecitation variance. Assuing (,..., ) (,..., ) are e p w w w q LP coefficients of clean speech noise wi p q Copright 7 ISCA 367

2 w w being e respective LP-odel orders. Let [,,, ], noise. w are e ecitation variances of clean speech The STP paraeter vector can be estiated b [5] ˆ NN N, Nw w i, j i j, ML, ML, ML, ML p(,, ' w,, w, ) p(, ) p( w, ) p( ) [ () ()... ( N )] T denotes e observed vector of nois saples for e current frae N is e frae leng. p( ) serves as a noralization ter can be obtained as () 3. MBE Fraewor Motivated b e high qualit speech snesized b MBE speech coding, we introduce e MBE AwS fraewor in is paper. In e MBE odel, speech spectru is divided into ultiple sub-bs each sub-b is sent for voicing judgent. Voiced sub-b is generated b periodic ecitation, unvoiced sub-b is generated b ro noise. In is paper, e enhanced spectru b codeboo-based eod is considered as an input paraeter for MBE odel. In general, e proposed MBE speech enhanceent fraewor can be divided into two stages, nael, speech analsis speech snesis. Fig. shows e proposed MBE AwS fraewor. The following will describe e fraewor in detail. N, Nw i j, ML, ML, ML, ML p( ) p(, w,,, w, ) p(, ) p( w, ) NN w i, j (3) N, Nw are e sizes of speech noise codeboo, respectivel. The odeled nois spectru is defined as S() Pitch Haronic Magnitude V/UV Decision Voiced Speech Snesizer Unvoiced Speech Snesizer Enhanced speech Pˆ ˆ / Aˆ ( ) ˆ / Aˆ ( ) (4) w w Speech Analsis Speech Snesis Aˆ ( ) Aˆ ( ) are e spectru corresponding to ˆ, ˆw, respectivel. The are given b w p q j ( ), ˆ j ( ) w w (5) Aˆ e A e Using e equivalence of e log-lielihood e Itaura- Saito distortion easure, we can obtain p(,,, ) Cep( d ( P, Pˆ )) (6) i j, ML, ML, ML w, w, IS C is a constant which has no influence on coputing STP paraeters. And e IS easure between e observed nois spectral envelope estiated nois spectral envelope is defined as ˆ P ( ) P ( ) dis ( P, P ) ( ln( ) ) d (7) Pˆ ( ) ˆ P( ) ' i j, ML, ML w, w, Besides, [,,, ] are e i entries,, ML, speech codeboo e, ML w,, i j j w noise codeboo are e aiu-lielihood estiates of speech noise ecitation variances which can be obtained b [4] C D (8) [ ] T w e atrices C D are given in [4]. Eventuall, e estiated AR coefficients can be used to construct a Wiener filter to obtain e enhanced speech ˆ / Aˆ ( ) H ( ) ˆ ˆ ˆ / ( ) ˆ A w/ Aw( ) (9) Figure : The proposed MBE fraewor. 3.. Speech analsis During e speech analsis stage, several acoustic paraeters are estiated including pitch period, haronic agnitude V/UV decision. First of all, e haronic agnitude corresponding to each cidate pitch period is calculated. The haronic agnitude is obtained b [8] A ( ) b ( ) a ( ) b ( ) S ( ) W(, ) a ( ) W(, ) () is a set of cidate pitch period. S ( ) is e enhanced spectru b e codeboo-based eod W(, ) is a periodic ecitation spectru wi period, which can be replaced b e frequenc response of e window function. Besides, a ( ) b ( ) are e upper lower bounds of e frequenc bins in haronic b. Pitch opt is obtained b iniizing e following error function ( ) [8] ( ) M ( ) M ( ) b ( ) a ( ) M ( ) [ S ( ) A ( ) W(, ) ] b ( ) a ( ) S ( ) is e nuber of haronics. After e optial pitch period opt haronic agnitude corresponding to autoaticall. () is fied, a set of opt is deterined 368

3 The last acoustic feature to be estiated in speech analsis stage is V/UV decision. When calculating e haronic agnitude pitch, it is assued at each haronic b is voiced. It will ae e snetic spectru in voiced portion close to e original spectru, while in unvoiced portion is ver different fro e original spectru. B using e difference a given reshold, we can get e V/UV decision. The difference function of haronic b D is defined as [8] D b ( ) opt opt a ( ) ˆ [ S ( ) S(, ) ] b ( ) opt opt a ( ) S ( ) opt () Sˆ ( opt, ) is e agnitude of reconstruct spectru, can be obtained b Sˆ (, ) A ( ) W(, ), a ( ) b ( ) (3) opt opt opt opt opt If D is lower an e given reshold T, e haronic b is considered as voiced, oerwise e haronic b is considered as unvoiced. In practice, T is set to. which can obtain a well perforance. db db frequenc(hz) (b) 6 Figure : An eaple of V/UV decision. (a) Codeboo-based eod enhanced spectru, (b) Ecitation spectru (c) V/UV decision. Fig. (a) (b) show e enhanced spectru of codeboo-based eod periodic ecitation spectru, respectivel. In Fig. (c), denotes voiced denotes unvoiced. As shown in e figure, if e atching error is sall, e haronic b is declared as voiced, oerwise is unvoiced. 3.. Speech snesis During e snesis stage, voiced unvoiced speech signals are snesized using different strategies. Each haronic of which is declared as voiced corresponds to a sinusoidal oscillator defined b e agnitude, frequenc phase. The haronic of voiced speech can be epressed as (a) frequenc(hz) (c) frequenc(hz) ( ) function of e s ( n) a ( n)cos[ ( n)], n N (4) v a n ( n ) are agnitude function phase haronic, respectivel. Here e phase function is obtained b sapling phase spectru of nois speech at each integer ultiples of pitch frequencies. B using e eod of tie doain snesis, it is possible to ae e snesized speech soo at e boundaries of e frae. The agnitude function is obtained b linear interpolation of each haronic agnitude ˆ n a ( ) ( ) [ ˆ () ˆ n A A A( )] (5) N Aˆ ( ) Aˆ () are haronic agnitude of previous current frae, respectivel. Siilarl, e pitch also needs to do a siilar linear interpolation between fraes. The final voiced speech is obtained b adding e sinusoidal function of each haronic M ( ) sˆ ( n) s ( n) (6) v For unvoiced speech, it is snesized b haronics which are declared as unvoiced. First, a ro Gaussian white noise is generated Fourier transfor is applied. Then, e spectru corresponding to e voiced sub-b coponents are set to zero. Finall, an inverse FFT is applied to get e final unvoiced speech. Fig. 3 shows e bloc diagra of unvoiced speech snesis. Gaussian White noise sequence Unvoiced speech v Windowed FFT IFFT Modif e spectru Figure 3: Unvoiced speech snesis. The enhanced speech is obtained b suing e voiced unvoiced speech. Because e phase spectru of e snesized speech is discrete, it is necessar to appl a FFT transfor replace e previous phase wi e original nois phase spectru Modified codeboo-based enhanced spectru In order to ae e estiation of acoustic features ore accurate, is paper cobines e codeboo-based Wiener filter wi speech presence probabilit before MBE fraewor. Let H ( ) denote at speech is present in frequenc bin ( )) denote e priori speech presence probabilit in frequenc bin. Since e codeboo-based eod is onl used to odel e spectral envelope, it is necessar to estiate e speech presence probabilit based on e frequenc bin. Here, Minia Controlled Recursive Averaging (MCRA) algori [9] is eploed to estiate e power spectru of noise, Pw ( ). Then e posteriori speech presence probabilit in frequenc bin can be estiated as [] ( )) ( ) ) ' ' ( )) ( )))( ( ))ep( ( )) (7) 369

4 ( ) ( )/ p( H ( )), ( ) ( ) ( )/( ( ) ) (8) ' ' ' ' ( ) denotes e priori SNR which is estiated b Decision-Directed (DD) eod [], ( ) denotes e posteriori SNR ( )) is set to.5. Above all, e codeboo-based Wiener filter can be iproved as ˆ / Aˆ ( ) ' H ( ) p( H( ) ) ˆ ˆ ˆ / ( ) ˆ A w/ Aw( ) 4. Results (9) In is section, we copare e perforance of proposed eod wi two traditional codeboo eods which coe fro [4] [5]. For coparison, we nae e as Ref. A Ref. B, respectivel. The test set is selected fro NTT database down-sapled to 8 Hz. It consists of nine utterances, four ales five feales. The speech is corrupted b babble, white, f6 factor noise fro NOISEX-9 database at db, 5dB db SNR. The size of saples per frae is 56 wi 5% overlapped. For codeboobased approach, a 5-bit speech codeboo 3-bit noise codeboos for each tpe of noise are trained. Besides, e order of AR odels for speech noise is. In order to deonstrate e effectiveness of e proposed eod, e coparison of spectrogras between e proposed eod e reference eods is presented in Fig. 4. (a) (b) (c) (d) Frequenc(Hz) Frequenc(Hz) Frequenc(Hz) Frequenc(Hz) Tie(s) Figure 4: Spectrogras of (a) nois speech (White noise db), (b) Ref. A, (c) Ref. B, (d) proposed. Fro e spectrogras, it is obvious at e proposed eod can generate clean haronics reove e noise between haronics. In addition, because e speech presence probabilit is used, e noise in silent segents can be furer reoved. The objective evaluation for speech enhanceent is perfored fro perceptual evaluation of speech qualit (PESQ) [], segental SNR (SSNR) [3] log-spectru distortion (LSD) [4]. The results are listed in Table, Table Table 3, respectivel. Table : Test Results of Average PESQ. Meod db 5dB db Nois Ref. A Ref. B proposed..5.7 Table : Test Results of Average SSNR Iproveent. Meod db 5dB db Ref. A Ref. B proposed Table 3: Test Results of Average LSD. Meod db 5dB db Nois Ref. A Ref. B proposed As can be seen fro Table, e PESQ score of e proposed eod is higher copared wi reference eods, especiall in low SNR conditions. Fro e results of SSNR LSD in Table Table 3, we can now at e proposed eod has less residual noise while has less spectru distortion. It is due to e AwS fraewor which can reconstruct clean haronic. In addition, because of e introduction of speech presence probabilit, e noise in silent segents can be furer reoved. Fro ree objective easures, we can now at e proposed MBE AwS fraewor has better effect in low SNR conditions. The reason is at e MBE odel is based on e etraction of acoustic paraeters, e reconstructed clean speech is obtained b ese paraeters. So in low SNR conditions, it can reove ore noise, especiall between e haronics. In e MBE speech analsis stage, we need to divide e frequenc b according to e haronics. It should be noted at a haronic b can have ore an one haronic. It is necessar to ae a coproise between e precise judgent of each haronic e continuit of e haronics. In e eperient a haronic b containing five haronics can get e best result. 5. Conclusions In is paper, we propose an iproved codeboo-based speech enhanceent based on MBE odel. It uses analsis-wisnesis fraewor to reconstruct speech rough acoustic features etracted fro nois speech. Copared wi traditional codeboo-based approach, it can reove noise not onl between adjacent haronics, but also in silent segents. The eperient results show a better perforance of e proposed eod copared wi reference eods under different noise conditions. 6. Acnowledgeents This wor was supported b e National Natural Science Foundation of China (Grant No. 6474, No. 635). 363

5 7. References [] S. F. Boll, Suppression of acoustic noise in speech using spectral subtraction, IEEE Transactions on Acoustics, Speech, Signal Processing, vol. 7, no., pp. 3-, 979. [] A. Aehrae, D. Pastor, A. Tataoui, Perceptual iproveent of Wiener filtering, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process., Las Vegas, pp. 8-84, 8. [3] R. Martin, Speech enhanceent based on iniu eansquare error estiation super-gaussian priors, IEEE Transactions on Speech Audio Processing, vol., no. 5, pp , Sep. 5. [4] S. Srinivasan, J. Sauelsson, W. B. Klen, Codeboo Driven Short-Ter Predictor Paraeter Estiation for Speech Enhanceent, IEEE Transactions on Audio, Speech, Language Processing, vol. 4, pp , Jan. 6. [5] S. Srinivasan, J. Sauelsson, W. B. Klen, Codeboobased Baesian speech enhanceent for nonstationar environents, IEEE Transactions on Audio, Speech, Language Processing, vol. 5, no., pp , Feb. 7. [6] P. Harding B. Milner, Speech enhanceent b reconstruction fro cleaned acoustic features, in Interspeech Proceedings,, pp [7] R. Chen, C.-F. Chan, H. C. So, Model-based speech enhanceent wi iproved spectral envelope estiation via dnaics tracing, IEEE Transactions on Audio, Speech, Language Processing, vol., no. 4, pp ,. [8] IEEE Transactions on Acoustics, Speech Signal Processing, vol. 36, no. 8, pp. 3 35, 988. [9] Cohen, I. Noise estiation b inia controlled recursive averaging for robust speech enhanceent, IEEE Signal Processing Letters, 9(), -5.. [] P. C. Loizou, Speech enhanceent: eor practice. Boca Raton, FL, USA: CRC Press, 7. [] Y. Ephrai D. Malah, Speech enhanceent using a iniu-ean square error short-tie spectral aplitude estiator, IEEE Transactions on Acoustics, Speech, Signal Processing, vol. ASSP-3, no. 6, pp. 9, Dec [] Perceptual Evaluation of Speech Qualit (PESQ), an Objective Meod for End-to-End Speech Qualit Assessent of Narrowb Telephone Networs Speech Codecs, ITU-T Recoendation, P.86, Feb,. [3] Quacenbush, S. R., Barnwell, T. P., Cleents, M. A., Objective Measures of Speech Qualit, Englewood Cliffs, NJ: Prentice Hall, 988. [4] Abrason, A., Cohen, I., Siultaneous Detection Estiation Approach for Speech Enhanceent, IEEE Transactions on Speech Audio Processing, 5(8), pp , 7 363