A Time-Frequency Adaptation Based on Quantum Neural Networks for Speech Enhancement

Transcription

1 A Tie-Frequeny Adaptation Based on Quantu Neural Networs for Speeh Enhaneent Kun-Ching Wang 1 and Chiun-Li Chin 1 Departent of Inforation Tehnology & Couniation Shin Chien University No. University Rd Neien Shiang Kaohsiung 845 Taiwan R.O.C w4@seed.net.tw Departent of Applied Inforation Sienes Chung Shan Medial University Taihung Taiwan R.O.C Abstrat: In this paper we propose a novel wavelet oeffiient threshold (WCT) depended on both tie and frequeny inforation for providing robustness to non-stationary and orrelated noisy environents. A pereptual wavelet filter-ban (PWFB) is firstly used to deopose the noisy speeh signal into ritial bands aording to ritial bands of psyho-aousti odel of huan auditory syste. The estiation of wavelet oeffiient threshold (WCT) is then adjusted with the posterior SNR whih is deterined by estiated noise power through the well-nown Quantu Neural Networs (QNN). In order to suppress the appearane of usial residual noise produed by thresholding proess we onsider asing properties of huan auditory syste to redue the effet of usial residual noise. Siulation results showed that the proposed syste is apable of reduing noise with little speeh degradation and the overall perforane is superior to several opetitive ethods. Key-Words: Speeh enhaneent; pereptual wavelet paet transforation; adaptive wavelet oeffiient threshold; usial residual noise. 1 Introdution Speeh enhaneent has beoe an iportant proble sine there are any areas where it is neessary to enhane the pereptual quality of speeh degraded by baground noise suh as voie ouniation and oding systes ar interiors for hands free ellular airraft opits hearing aids and autoati speeh reognition (ASR) systes [1- ]. So far the researhers provide any approahes to enhane speeh quality [3-1]. Wavelet thresholding is a siple de-noise tehnique that adequately hooses the value of wavelet oeffiient threshold (WCT) to reove noise for signal in any signal-proessing appliations [7-1]. Donoho and Johnston [7-8] proposed a universal threshold for reoving the additive white Gaussian noise but ay not wor well in enhaning olored-noise orrupted signal. After that adaptive wavelet-based ethods in speeh enhaneent are widely presented [9-11]. They utilize variant WCT to iprove the perforane of speeh enhaneent. Bahoura et al. [11] proposed a ethod of threshold adaptation in tie doain. Utilizing the use of Teager energy operator (TEO) to iprove the disriinability for a speeh frae whether is speeh-doinated or noise-doinated. In fat the ey issue is to selet threshold and shrinage funtion in wavelet thresholding ethod. Traditional wavelet de-noising ethods involve either hard or soft thresholding. In hard thresholding ethod the oeffiient is set to a speifi value when its agnitude exeeds the threshold. On the other hand soft thresholding shrins or sales the oeffiient that exeeds the threshold value. It is nown that Quantu Neural Networs (QNN) is exploited a new shrinage funtion in speeh enhaneent syste [1]. To redue the effet of usial residual noise a nuber of ethods were onsidered [13-14]. Virag [13] ade use of asing properties of the huan auditory syste to redue the effet of residual noise. Sine huan ears annot pereive additive noise when at levels below the noise asing threshold (NMT). In this paper we propose a novel WCT depended on both tie and frequeny inforation for providing robustness to non-stationary and orrelated noisy environents. A pereptual wavelet filter-ban (PWFB) is firstly used to deopose the noisy speeh signal into ritial bands aording to ritial bands of psyho-aousti odel of huan auditory syste. The estiation of WCT is then adjusted with the posterior SNR whih is deterined by estiated noise power through the well-nown QNN. In order to suppress the appearane of usial residual noise produed by thresholding proess we onsider asing properties of huan auditory syste to redue the effet of usial residual noise. ISSN: Issue 1 Volue 7 January 1

2 Pereptual Wavelet Filter-ban The huan speeh ostly spans within 4 Hz and there are only 17 ritial bands existed in this bandwidth as listed in Table 1 [13]. In order to introdue a speeh enhaneent ethod based on the huan auditory odel a pereptual wavelet filter-ban (PWFB) is designed to ii the ritial bands as widely used in pereptual [15]. The filter bans ipleented by using the high-pass filter and low-pass filter with the Daubehies faily wavelet [16] pereptually divide whole band into subband doain. In the first level deoposition saling spae and wavelet spae will be deoposed into two whih orrespond to the frequeny ranges of and 4 Hz. This operation is repeated to at ost five ties. Table shows the oeffiients through five-stage tree struture of pereptual wavelet filter-ban (PWFB). The PWFB deoposes the noisy signal x( n ) into 17-subbands orresponding to wavelet oeffiient sets w = PWFB{ x( n) } n = 1... N (1) where PWFB{} eans the pereptual wavelet paet transfor. w defines the th wavelet oeffiient in th subband. N is the length of speeh frae. 3 The Estiation of Tie-Frequeny Dependent WCT Here we use a new adaptive tie-frequeny dependent thresholds estiation ethod. This involves first estiating the standard deviation of the noise σ for every subband and tie frae. Consequently we use a quantile-based noise traing approah to tra the slowly varying nonstationary noise statistis [17]. Table 1. The harateristis of ritial bands under 4 Hz Bar-Band Nuber Lower Edge Upper Edge Center frequeny Bandwidth (Hz) (Hz) (Hz) (Hz) Table. The oeffiients of pereptual wavelet filterbans Bar-Band Nuber Transfor stage Coeffiients index Coeffiients length The noise level estiation is given by int( q L seg ) = w q Lseg j= ɶ σ int( ) () where int( ) is the nearest integer rounding funtion. The noinal value of q is.. σɶ is denoted as the orresponding estiated noise level of the th frae in the th subband. These are estiated using the segent of previous data w =... L 1. { seq } where L seq eans the length of the segent in the th subband. The initial WCT WCT for th subband at the th frae an be estiated as in [17]: WCT = ɶ σ log( L log ( L )) (3) fr fr where L fr eans the frae length at the th subband. The posteriori SNR on th subband an be evaluated as w SNRpot = 1 log 1 ɶ σ (4) where σɶ and w denote the estiated subband noise power and observed signal power. To reedy the drawbas of the traditional threshold algoriths we adopt a new ethod of S- urve of Quantu Neural Networ (QNN) to hoie appropriate WCT [18]. The S-urse in QNN odel is ulti-level whih expression is n 1 s 1 QNN( x) = (5) n 1+ exp( ( x θ )) s i= 1 i ISSN: Issue 1 Volue 7 January 1

3 where n s is the nuber of urve level and θ i is the position of the level respetively. x is herein defined as x = α ( SNR T ). pot 4 Pereptual Suppression using Noise Masing Threshold (NMT) In order to iprove the final pereptual quality a suppression ethod of usial residual noise an adopt a pereptual gain fator into wavelet thresholding. The tie-frequeny adapted wavelet threshold is finally odified as below: WCT = WCT Gain where final pep (6) ɶ σ Gainpep = 1 1+ ax 1 NMT ( ) denotes a pereptual gain fator given by [17]. Fro Eq.(6) it is nown that if the energy of usial residual noise σɶ is greater than the NMT NMT ( ) in a subband the wavelet oeffiient thresholds beoe sall adjusted by the gain fator to suppress infeting noise. However if the energy of residual noise is saller than the NMT the orrupting noise annot be pereived by the huan ear. We do not need to hange the WCTs for retaining the speeh quality. In order to alulate the NMT on eah subband the estiated spetra of enhaned speeh ust be first deterined and it be roughly estiated by the spetral-subtration ethod. Netx the subband energy ε ( ) is alulated by h ε ( ) = w (7) l where h and l denote the upper and the lower frequenies at ritial band an be found in [13]. An exitation pattern B( ) an be regarded as an energy distribution along the basilar ebrane. B( ) an be alulated by onvolving the subband energy ε ( ) with the spreading funtion F( ). B( ) is given by [13 19]: B( ) = F( ) ε ( ). (8) A relative threshold offset O( ) whih an be found in [13 19] speifies whether a speeh frae is tone-lie or noise-lie. This threshold should be iposed when adjusting the log subband energy. Therefore a threshold Bɶ ( ) is oputed as the su of the log energy for the exitation pattern and the offset O( ) written as Bɶ ( ) = 1 log 1 B( ) + O( ) (9) where the values of the offset O( ) are all negative. Convolving the subband energy ε ( ) with the spreading funtion F( ) inreases the energy in eah subband so to ultiply eah Bɶ ( ) by the inverse of the energy gain is neessary for renoralization. Aordingly a noralized threshold is given by Th( ) = Bɶ ( ) G( ) (1) where G( ) denotes the gain fator between the spread energy B( ) and the subband energy ε ( ) in db. G( ) is expressed as B( ) G( ) = 1 log1. (11) ε ( ) Additionally the noralized threshold Th( ) is opared with the absolute-hearing threshold (AHT) whih is frequeny-dependent and an be losely approxiated as [13 19].8.6( f 3.3) AHT( f ) = 3.64 f 6.5e f [db] with f in ilohertz. Finally the NMT NMT ( ) is obtained by { } (1) NMT ( ) = ax AHT( f ) Th( ) (13) where f is hosen as the entral frequeny of the ritial band. 5 Ipleentation of the Proposed Algorith Figure shows the syste blo diagra for the proposed wavelet-based speeh enhaneent algorith. The wavelet paet filter-ban is first applied to deopose the noisy speeh signal into ulti-resolution tie-spetral subbands. Thresholds are independently estiated aross suessive speeh fraes in eah deoposed subband and are adapted as tie-variant values based on the adaptive noise estiation algorith. The suppression of baground noise is then ahieved by soft thresholding the deoposed wavelet oeffiients. Finally these thresholded wavelet oeffiients are reonstruted to obtain the enhaned speeh saples using the inverse wavelet paet filter-ban. ISSN: Issue 1 Volue 7 January 1

4 Figure. The arhiteture of proposed speeh enhaneent ethod based on the tie-frequeny adaptation of the wavelet threshold 3.1 Segent SNR Iproveent The average SegSNR an be used to estiate the Table 3. The average SegSNR results of speeh aount of noise redution residual and speeh enhaneent under various noisy onditions distortion. Table 3 shows that the average SegSNR Noise type average SegSNR (db) results of the speeh enhaneent evaluations in Proposed S.H. Chen M. Bahoura different SNR levels. Fro this Table we an see F16 noise that the proposed algorith has uh better enhaneent perforane than others. White noise Babble noise Fatory noise Vehile noise Table 4. MOS results of the listening test The ean opinion sore Noise type Proposed S.H. Chen M. Bahoura F16 noise White noise Babble noise Fatory noise Vehile noise Experiental Results In this experients the speeh databases are Mandarin and spoen by ales and 1 feales. The frae size is 56 at the sapling rate of 8 Hz with 16-bit resolution. Noisy speeh signals were obtained by orrupted the lean speeh with White F16 and Babble (speeh-lie) noises extrated fro the Noisex-9 database []. The reorded speeh signal was tested in the noisy environent inluding SNRs range fro - 5dB to 1 db. In our experients the subjetive evaluation and objetive evaluation are applied to evaluate the perforane of the speeh enhaneent ethod. 3. Subjetive Listening Tests The ean opinion sore (MOS) [1] was used to represent the global pereption of the residual noise baground noise and speeh distortion. The MOS is subjetively evaluated the subjetive listening tests by a five-sale absolute opinion fro 1 (poor) to 5 (exellent). The results of subjetive listening tests are presented in Table 4. The subjetive listening tests show that the proposed enhaneent ethod produes the highest quality speeh pereived by the atual huan listeners aong the algoriths being tested espeially for low SNR. The pereptual ethod is better able to reove the baground noise than that without pereptual odel. 7 Conlusion In the paper a novel speeh enhaneent algorith using tie-frequeny wavelet threshold is presented. In order to exploit the physiology of huan auditory syste to reovery high-quality speeh fro noisy speeh the noisy speeh is first deoposed into ritial bands by pereptual wavelet paet transfor. Then an adaptive wavelet threshold is adjusted aording to posterior SNR based on S- urve of Quantu Neural Networ (QNN). Experiental results show that the proposed algorith is better able to pereptually redue the non-stationary and olored noise and is free fro usial residual noise. ISSN: Issue 1 Volue 7 January 1

5 8 Anowledgent This researh was partially sponsored by the National Siene Counil Taiwan under ontrat nuber NSC 98-1-E Referenes: [1] B.H. Juang Reent Developents in Speeh Reognition under Adverse Conditions in Proeedings of Int. Conf. Spoen Language Proess '9 199 pp [] J.H. Chen and A. Gersho Adaptive Postfiltering for Quality Enhaneent of Coded Speeh IEEE Trans. Speeh and Audio Proessing Vol pp [3] S.F. Boll Suppression of aousti noise in speeh using spetral subtration IEEE Trans. Aoustis Speeh Signal Proess Vol pp [4] J.R. Deller J.H.L. Hansen and J.G. Proais Disrete-Tie Proessing of Speeh Signals seond ed. IEEE Press New Yor. [5] S. Hayin Adaptive Filter Theory third ed. Prentie Hall Upper Saddle River New Jersey [6] Y. Ephrai and D. Malah Speeh enhaneent using a iniu ean-square error short-tie spetral aplitude estiator IEEE Trans. Aoust. Speeh Signal Proess Vol. ASSP-3. No pp [7] D.L. Donoho De-noising by Soft Thresholding IEEE Trans. Infor. Theory Vol. 41 May 1995 pp [8] D.L. Donoho and I.M. Johnstone Ideal Spatial Adaptation by Wavelet Shrinage Bioetria Vol. 81 No pp [9] S.H. Chen and J.F. Wang Speeh Enhaneent Using Pereptual Wavelet Paet Deoposition and Teager Energy Operator Journal of VLSI Signal Proessing Vol pp [1] S.F. Lei and Y.K. Tung Speeh Enhaneent for Nonstationary Noises by Wavelet Paet Transfor and Adaptive Noise Estiation Proeedings of 5 International Syposiu on Intelligent Signal Proessing and Couniation Systes De. 5 pp [11] M. Bahoura and J. Rouat Wavelet speeh enhaneent based on the teager energy operator IEEE Signal Proess. Lett. Vol. 8 No. 1 1 pp [1] L Fei Z Shengei and Z Baoyu Quantu neural networ in speeh reognition - Pro. IEEE Int. Conf. Signal Proess [13] N. Virag Single hannel speeh enhaneent based on asing properties of the huan auditory syste IEEE Trans. Speeh Audio Proess. Vol. 7 No. Mar pp [14] Y. Hu and P.C. Loizou Speeh enhaneent based on wavelet thresholding the ultitaper spetru IEEE Trans. Speeh Audio Proess Vol. 1 No. 1 4 pp [15] E. Zwier and H. Fastl Psyhoaoustis: Fats and Models Springer-Verlag New Yor 199. [16] S. Mallat Multifrequeny hannel deoposition of iages and wavelet odel IEEE Trans. Aoust. Speeh Signal Proess Vol pp [17] Q Fu and E.A. Wan Pereptual wavelet adaptive denoising of speeh Eighth European Conferene on Speeh 3 [18] J. F. Teng J. Dong S. Y. Wang H. Bao M. G. Wang A Speeh Enhaneent Algorith Based on Bar-Sale Wavelet Paage proeedings of the sixth international onferene on ahine learning and ybernetis Hong Kong August 7 pp.19-. [19] C.T. Lu and H.C. Wang Speeh enhaneent using pereptually onstrained gain fators in ritial-band-wavelet paet transfor IEE Eletron. Lett. Vol. 4 No. 6 4 pp [] Varga and H.J.M. Steeneen Assessent for autoati speeh reognition: II. NOISEX-9: A database and an experient to study the effet of additive noise on speeh reognition systes Speeh Coun. Vol pp [1] J. Deller J. Proais and J. Hansen Disrete- Tie Proessing of Speeh Signals Englewood Cliffs NJ: Prentie-Hall ISSN: Issue 1 Volue 7 January 1