Speech Enhancement Based on Analysis Synthesis Framework With Improved Pitch Estimation and Spectral Envelope Enhancement

Transcription

1 ICSP014 Proceedngs Speech Enhancement Based on Analyss Synthess Framework Wth Improved Ptch Estmaton and Spectral Envelope Enhancement Bn Lu 1, Fuyuan Mo, Janhua Tao 1 1 atonal Laboratory of Pattern Recognton, Insttute of Automaton, Chnese Academy of Scences, Bejng Insttute of Acoustcs, Chnese Academy of Scences, Bejng lubn@nlpr.a.ac.cn, mofuyuan@alyun.com, jhtao@nlpr.a.ac.cn, Abstract Ths paper presents a speech enhancement approach based on analyss synthess framework. An mproved mult-band summary correlogram (MBSC) algorthm s proposed for ptch estmaton and voced/unvoced (V/UV) detecton. The proposed ptch detecton algorthm acheves a lower ptch detecton error compared wth the reference algorthm. The denosng autoencoder (DAE) s appled to enhance the lne spectrum frequences (LSFs). The reconstructon loss could be decreased compare wth the swallow model. The proposed approach s evaluated usng the perceptual evaluaton of speech qualty (PESQ) and the expermental results show that the proposed approach mproves the performance of speech enhancement compared wth the conventonal speech enhancement approach. In addton, t could be appled to parametrc speech codng even at low bt rate and low SR envronments. Keywords- analyss-synthess framework, mult-band summary correlogram, denosng autoencoder, speech enhancement, speech codng. I. ITRODUCTIO Sngle-channel speech enhancement s an mportant branch of speech sgnal processng. It s useful for many applcatons such as speech recognton, speech codng and hearng ad; t s often used as a pre-processer to mprove speech qualty and ntellgence. Enhancng the speech sgnal from corrupted sgnal had been addressed as a challengng topc. It s partcularly dffcult to track the non-statonary especally n low SR envronments from a sngle -channel nosy speech. A large number of speech enhancement approaches are already proposed such as spectral subtracton [1] and mnmum mean square error (MMSE) []. A major drawback of the spectral subtracton approach s the ntroduced muscal n the enhanced speech. The MMSE methods attract a great deal of nterest and generally outperform algorthms n other categores n varous nosy condtons. The man reason s twofold. Frst, they are optmzed n a best spectral magntude estmaton by notcng the unmportance of the phase n speech enhancement. Second, they take advantage of a pror knowledge estmated usng a Bayesan framework. However, the MMSE algorthms often have poor performance n nonstatonary or low SR envronments. In [3], an analyss synthess framework s proposed to resynthesze clean speech sgnals based on acoustc parameters (ptch, spectral gan and spectral envelope) extracted from nosy speech. The target speech s reconstructed wth related acoustc parameters only and background s automatcally removed. Ths approach s attractve due to t can retreve damaged harmonc structure and at the same tme elmnates muscal. However, to ensure accurate model parameter estmaton, a preprocessng step s often requred to preclean the nosy sgnals pror to speech enhancement based on analyss synthess framework. It s reported n [3] that ptch and spectral gan estmaton appled on precleaned spectrum can gve satsfactory result even n very low SR envronments. However, most precleanng algorthms dffcult to recover the spectral envelope whch has been dstorted by background. Both ptch estmaton and spectrum envelope estmaton are mportant to mprovng the performance for speech enhancement. Ptch estmaton algorthms can be broadly classfed nto three categores: tme-doman, frequency-doman, and tmefrequency-doman. Tme-doman F0 estmaton drectly explot a sgnal s temporal perodcty [4]. Frequency-doman F0 estmaton make use of the sgnal s short-tme spectral harmoncty [5]. Tme-frequency doman F0 estmaton often separates a sgnal nto varous frequency bands, and then apples tme-doman processng n each frequency band [6]. The audtory-model correlogram-based F0 estmaton s a popular tme-frequency doman method. It can yeld estmates close to human s perceved ptch for sgnals and also have the potental to be -robust. An SR-weghted correlogrambased F0 estmaton usng mult-band comb FBKs s proposed n [7]. The proposed F0 estmaton s effectve n mprovng the accuracy of F0 estmaton n the presence of. The subband whch has hgh vocng strength represents obvous harmonc structure and t s effectve to ptch estmaton. The method to mprove the spectral envelope estmaton can be regarded as a problem of estmatng spectral envelope parameters of clean speech from nosy speech. It s well known that Wener flterng s correlated wth lnear predcton, and clean spectral envelope parameters can be teratvely estmated from nosy speech usng Wener flterng [8]. Besdes, the Kalman flter s also wdely studed n speech enhancement. In [9], t ncorporates Kalman flter to track the temporal trajectores of lne spectrum frequences (LSFs). The enhanced LSFs are then drected nto the analyss synthess framework to mprove the spectral envelope estmaton, and hence the performance of speech enhancement. The Gaussan mxture model (GMM) s a typcal mappng model whch s wdely appled to voce converson [10] and artfcal wdeband /14/$ IEEE 461

2 extenson [11]. It s effectve to reconstruct the target spectrum envelope. Deep learnng has emerged as a new area of machne learnng research [1]. It can dscover the underlyng regularty of multple features, and have strong generalzaton abltes than shallow models. The basc strategy s to tran a deep network wth greedy layer wsed pre-tranng plus fne tunng. For spectral envelope enhancement, we are payng attenton to learnng a mappng between the nosy spectral envelope and the clean spectral envelope. The denosng autoencoder (DAE) s an equvalent model [13]. It s traned to reconstruct the orgnal spectral envelope from a corrupted spectral envelope. It s a mult-layer neural network structure. The model s traned to mnmze the reconstructon loss between the orgnal spectral envelope and the reconstructed spectral envelope. In ths paper, we present a speech enhancement approach based on analyss synthess framework to enhance nosy speech sgnals. A prelmnary precleanng step s requred to preclean the nosy sgnals. The goal of the precleanng step s to flter the nosy sgnals such that t s more sutable for the analyss synthess framework. An mproved mult-band summary correlogram (MBSC) ptch detecton algorthm s proposed. Ths work s an extenson of the algorthm proposed n [14]. To mprove the -robustness of V/UV detecton and ptch estmaton, the subband whch has the hgh vocng strength s selected and the lnear predcton resdual sgnal s consdered. The denosng autoencoder (DAE) s appled to buld the mappng relatonshp between pre-cleaned LSFs and clean LSFs. The enhanced LSFs are then drected nto the analyss synthess framework to mprove the spectral envelope estmaton. The proposed algorthm takes advantage of the analyss synthess framework to effectvely elmnate muscal. On the other hand, t looks for the mappng relatonshp to obtan enhanced spectral envelope through deep layer network. The could be suppressed and the dstorted speech could be restored effectvely. The rest of ths paper s organzed as follows: the detal of proposed approach s ntroduced n secton. The evaluaton results are showed n the secton 3. Concluson wll be elaborated n the secton 4. the precleanng stage, the logmmse s appled to estmate the short-tme spectral ampltude for each frame. The mproved MBSC s appled to both ptch estmaton and voced/unvoced detecton n dfferent subband. We select the DAE model to enhance the LSFs. The spectral gan estmaton s measured usng a ptch adaptve wndow length. All estmated and enhanced speech parameters are sent to the syntheszer based on mxed exctng vocoder to reconstruct the target speech. Detaled procedures for proposed algorthm are shown as follows. A. Pre-cleanng The logmmse s appled to precleanng. The estmaton of the short-tme spectral ampltude (STSA) s formulate as that of estmatng the ampltude of DFT coeffcents of the speech sgnal x(t), gven the nosy observatons y(t). The DFT coeffcents of the speech sgnal, as well as of the sgnal are modeled as statstcally ndependent Gaussan random j k varables. Let X A e α j k =, D, and Y = R e β denote the kth k k k k k DFT coeffcent of the speech sgnal, the sgnal, and the nosy observatons, respectvely. We are lookng for the estmator A ^, whch mnmzed the log ampltude spectrum k dstorton. Ths method s superor to the MMSE STSA estmator snce t results n a much lower resdual level wthout further affectng the speech tself [15]. The nosy speech s ntally processed by logmmse to estmate clean speech and the speech sgnal s reconstructed approxmately. There wll be large dstorton n pre-cleaned speech especally for nonstatonary. It s mportant to restore the dstorted speech. B. Ptch estmaton and voced/unvoced detecton The block dagram n Fg. 1 gves an overvew of the proposed MBSC ptch detector. Ths s an extenson of the ptch estmaton algorthm proposed n [14] II. PROPOSED ALGORITHM In ths secton, we frstly ntroduce the framework of the proposed speech enhancement approach. Subsequently, the further detals are presented. The flowchart of proposed algorthm s shown n Fg. 1. Fgure 1: Block dagram of the proposed algorthm There are total sx parts n the proposed algorthm whch ncludes precleanng, ptch estmaton, V/UV detecton, LSFs enhancement, spectral gan estmaton and speech synthess. In Fgure : Block dagram of the proposed ptch detector The nput speech sgnal s frst decomposed nto four subbands usng 3-pont FIR flters. A 1-kHz flter bandwdth 46

3 s chosen so that at least two harmoncs are captured by each flter. The Hlbert envelope n each subband s extracted. We consder the lnear predcton resdual from the frst subband n subsequent processng. It s more effectve for ptch detecton, especally when the frst harmonc s not attenuated or corrupted compare wth the orgnal sgnal. Mult-channel comb flterng s performed separately for each subband. Ths comb-functon enhances spectral harmoncs and suppresses the energes at the subharmoncs. Harmonc-to-subharmonc energy rato (HSR) s a measure computed to ad the selecton of relable comb-flter channels for each subband. The HSR of the kth channel n subband s, denoted by q s computed usng Eq. (1). st, (k) (1) q (k) = X (f)c (f) / X (f)(1 c (f)) st, kst,, k kst,, k f f where X (f) and c (f) denote the DFT coeffcents n k, s, t k frame t for nput speech and comb-flter respectvely. Channel selecton s performed on each subband, usng a three stages selecton process desgned to mprove both ptch estmaton and vocng detecton performance. The channel s selected accordng to q, (k) and autocorrelaton (ACR). The st detal nformaton for channel selecton s formulated n [14]. After channel selecton, an HSR-based weghted averagng scheme s performed. Through ths weghtng scheme, ACR from the more relable channels wll have a greater mpact on ther subband, resultng n a more promnent ACR peak at the most lkely ptch perod of the sgnal. To mprove the -robustness of V/UV detecton and ptch estmaton, we select the subband whch has the hgh vocng strength. The vocng strength s confrmed based on selected channel. A constant threshold s appled on the maxmum nterpolated peak ampltude to obtan the V/UV decson for correspondng subband. The subband selecton s mplemented for three subbands (1-k, -3k and 3-4k). The frst subband (0-1k) s more robust for ptch detecton and s always selected. It s effectve to elmnate nterfere of perodcty. The subband-relablty-weghtng scheme s performed n selected subband. It could reduce the varablty of the maxmum peak ampltude n the MBSC for nosy speech. The ptch canddates correspondng to the 10 hghest peaks are dentfed. Each peak and ts mmedate neghbors are ftted by a parabola, and the ampltude and lag poston correspondng to the maxmum pont of ths parabola are the refned ptch measurements for the respectve ptch canddate. As for V/UV detecton, a constant threshold s appled on the maxmum nterpolated peak ampltude to obtan the ntal V/UV decson for each frame. Ths s followed by a 5-pont medan flterng n tme on these ntal decsons to get the fnal V/UV detecton. C. LSFs enhancment The flowchart of LSFs enhancement s shown n Fg. 3. Fgure 3: Block dagram of the proposed LSFs enhancement In ths subsecton, the DAE s appled to LSFs enhancement. Assumng that X and Y are the normalzed LSFs of the clean speech and pre-cleaned speech respectvely for each frames. The LSFs are normalzed accordng to mean and varable for each dmenson. We can formulate the LSFs enhancement process as X = η( Y ) () whereη s the LSFs enhancement process that reconstructs the LSFs based on DAE. The objectve of LSFs enhancement can be expressed as: ϕ = arg mn E [ η(y ) X ] (3) x The task s to fnd the ϕ that s the best estmaton of η. We use the speech par (pre-cleaned LSFs and clean LSFs) to tran the DAE. For each hdden layer neural autoencoder, t ncludes one nonlnear encodng stage and one lnear decodng stage as: h(y ) = σ (W 1y+ b) x = Wh(y ) + c where W and W are encodng and decodng matrx as the 1 neural network connecton weghts. W 1 = W T s used as one regulaton. b and c are the bases vectors of nput and output 1 layer, respectvely. σ ( x ) = (1 + exp( x )) s the sgmod actvaton functon and h s the actvaton of hdden layer. In the proposed algorthm, a mult hdden layers autoencoder wll be traned wth the pre-cleaned LSFs as nput and the clean LSFs as output. We adopt greedy layer wsed pre-tranng plus fne tunng to tran the DAE. The tranng par for the frst DAE s X and Y. Then the tranng par for the next DAE wll be h( X ) and h( Y ). After pre-tranng of each layer, all the layers are stacked to form a deep autoencoder for fne tunng. In the fne tunng stage, the ntal network parameters are fxed as the parameter obtaned from pretranng stage. The parameter s adjusted based on back propagaton used n neural network. The DAE s effectve to restore the dstorted spectrum and enhance LSFs due to the deep models have strong generalzaton abltes than the shallow models. D. Gan estmaton The nput speech sgnal gan s measured twce per frame usng a ptch adaptve wndow length [16]. Ths length s dentcal for both gan measurements and s determned as follows. For (4) 463

4 voced frame, the wndow length s the shortest multple of estmated ptch whch s longer than 10 samples. For unvoced frame the wndow length s 10 samples. The gan calculaton for the frst wndow produces and s centered 80 samples before the last sample n the current frame. The calculaton for the second wndow produces and s centered on the last sample n the current frame. For proposed algorthm, the frame length s 160 samples. The gan s the RMS value, measured n db, of the sgnal n the wndow, s n. where L s the wndow length. 1 G = + (5) L 10 log (0.01 s ) 10 n L n = 1 E. Speech synthess based on mxed exctng vocoder The mxed exctaton s mplemented usng a mult-band mxng model. The effect of ths mxed exctaton s to reduce the buzz usually assocated wth LPC vocoders. The relatve pulse and power n each frequency band s determned by an estmate of the vocng strength at that frequency n the nput speech. The vocng strength for dfferent subband s confrmed accordng to proposed ptch estmaton algorthm n ths paper. The mxed exctaton s generated as the sum of the fltered pulse and exctatons. The pulse flter for the current frame s gven by the sum of all the bandpass flter coeffcents for the voced frequency bands, whle the flter s gven by the sum of the bandpass flter coeffcents for the unvoced bands. The adaptve spectral enhancement flter s appled to the mxed exctaton sgnal. Ths flter s a tenth order pole/zero flters wth addtonal frst-order tlt compensaton. Its coeffcents are generated by bandwdth expanson of the lnear predcton flter transfer functon correspondng to the nterpolated LSF s. Snce the exctaton s generated at an arbtrary level, the speech gan must be ntroduced to the syntheszed speech. The correct scalng factor s computed for each syntheszed ptch perod of length. The pulse dsperson flter s a 65th order FIR flter derved from a spectrally flattened trangle pulse. It could reduce the harsh ngredents of syntheszed speech. III. EXPERIMETS AD RESULT AALYSIS A. Data and methodology In ths secton, we evaluate the proposed approach on ptch estmaton and V/UV decson, LSFs enhancement, speech enhancement based on analyss-synthess framework, parametrc speech codng at low bt rate. In ths test, the clean speech samples are selected from TIMIT database [17]. Four types of recordngs extracted from the osex-9 database [18], namely pnk, factory, volvo and buccaneer, were used as the sgnals. Three SR condtons, 0dB, 5dB and 10dB, are ncluded n the tranng process. The speech sgnal s down-sampled to 8KHz. For speech enhancement, the frame length s 160 samples and the frame shft s 80 samples. For speech codng, the frame length s the same as MELP standard. The 3000 utterances selected randomly from the tranng set of the TIMIT database were added wth the above mentoned four types of and three levels of SR. Another 300 randomly selected utterances from the TIMIT database were used to construct the test set for each combnaton of types and SR levels. In addton, we select another 500 randomly selected utterances from the TIMIT database to confrm both the DAE parameter and GMM parameter. Two other types, namely whte and babble were used for msmatch evaluaton. For ptch estmaton and V/UV decson, three types of error metrcs are commonly used. The frst s Vocng Decson Error (VDE). The second s F0 value estmaton error called the Gross Ptch Error (GPE). The FFE takes both GPE and VDE nto consderaton. [19] + V U U V VDE = *100% (6) = (7) F0 E GPE *100% VV VV FFE = * GPE + VDE (8) For LSFs enhancement, the reference algorthm s Gaussan Mxture Model (GMM) whch s used wdely n the LSFs transformaton. We evaluate the performance of proposed method wth the dstance between the reconstructed LSFs and the target LSFs accordng to (9). 10 d( lsf, lsf ) = å lsf -lsf (9) s t s t = 1 For speech enhancement, the reference algorthm s logmmse [15] and subspace []. We evaluate the performance of proposed method wth the perceptual evaluaton of speech qualty (PESQ) [0]. The PESQ, whch s a mean opnon score, s also used to evaluate the qualty of the restored speech. It has better correlaton wth subjectve tests than the other objectve measures. In addton, we evaluate the proposed approach n low bte rate speech codng system. The mxed exctaton lnear predcton (MELP) s the most mature parametrc speech codng method so far. Therefore, we select MELP standard to evaluate the proposed algorthm. PESQ s also used for the performance evaluaton at low bt rate speech codng. B. Parameter determnaton In ths secton, we descrbe the experments to choose the optmal parameter for DAE model and GMM. For DAE model, we search over a range of parameter to confrm the number of hdden unts (30, 50, 70 and 90) and the number of hdden layers rangng from 1 to 3. For GMM, we search the number of gaussan dstrbuton (16, 3, 64 and 18). In ths study, we set the archtecture of a DAE as follows: n encodng stage, the sze of nput layer s 10, each hdden layer s 50. All the layers are stacked and unrolled to form a deep autoencoder layer szes are In our experments, a batch sze of 100 was used. The number of epoch for each layer of pre-tranng was 0. And n fne tunng stage, the maxmum number of 464

5 teraton was set to 100. We optmze the learnng rate rangng from to 0.05 (the step s 0.005). The learnng rate was set at 0.0. We confrm the number of gaussan dstrbuton s 64 for GMM. C. The evaluaton for ptch estmaton and V/UV decson Three other ptch estmaton algorthms are also evaluated. The three reference algorthms nclude: GetF0 [1], MELP [16], and MBSC proposed n [14] whch the subband selecton and lnear predcton resdual are not consdered. The reference ptch values have been obtaned automatcally and thoroughly revsed manually n the way descrbed n [3]. The performance of the ptch estmaton and V/UV decson for dfferent method and dfferent confgure as shown n Table 1 and Table respectvely. Table1 evaluate the performance n 0dB condton and Table tabulates the performance, averaged dfferent SR (0dB, 5dB, 10dB). Table 1. GPE, VDE and FFE of PEAs (SR=0dB) ose type Method VDE(%) GPE(%) FFE(%) Whte Volvo GetF MELP MBSC Proposed GetF MELP MBSC Proposed GetF MELP MBSC Proposed Table. GPE, VDE and FFE of PEAs (averaged dfferent SR) ose type Method VDE(%) GPE(%) FFE(%) Whte Volvo GetF MELP MBSC Proposed GetF MELP MBSC Proposed GetF MELP MBSC Proposed Table 1 and Table compare performance of the evaluated algorthms on dfferent types of. In general, proposed algorthm gves the lowest GPE and FFE n dfferent types of. For wde-band whch nclude babble and whte. The MBSC proposed n [14] s pror to the proposed algorthm slghtly n terms of VDE. For narrow-band whch ncludes Volvo, proposed method has lowest rate n terms of VDE. The advantage of proposed algorthm s due to the subband whch has the hgh vocng strength s selected and the lnear predcton resdual sgnal s consdered n the process of ptch estmaton, whch can effectvely attenuate the especally for narrow-band. D. The evaluaton for LSFs enhancement In ths subsecton, we evaluate the LSFs reconstructed error for dfferent approach whch ncludes GMM and DAE. In ths evaluaton, dfferent types of (pnk, factory, buccaneer and volvo) for match evaluaton; whte and babble for msmatch evaluaton) s consdered. The results of reconstructed error are shown n Fg.4. From ths fgure, we can see that DAE s more effectve compared wth GMM for both the match evaluaton and the msmatch evaluaton. The proposed method s more robust n dfferent nosy envronment even not nclude n tranng set. The precleaned LSFs could be enhanced effectvely through DAE. It s due to the deep models have strong generalzaton abltes than the shallow models. Fgure 4: The average LSFs reconstructed error E. The evaluaton for speech enhancement The proposed method based on analyss-synthess framework s compared wth two dfferent methods, ncludng logmmse and subspace. The degraded speech wthout enhancement s denoted as nosy. The PESQ scores are shown n Table.3. Table 3 the PESQ results for speech enhancement ose type Method 0dB 5dB 10dB Match osy Logmmse Subspace Whte Proposed osy Logmmse Subspace Proposed osy Logmmse Subspace Proposed The PESQ of match s the average PESQ of four type whch nclude pnk, factory, buccaneer and volvo. As shown n Table3, we can see that the proposed method s more effectve compared wth the dfferent reference methods. It s noted that an average of around 0.3-pont mprovement over the best conventonal method are acheved n varous condtons. In comparson wth the dfferent reference methods, the proposed method acheves better objectve speech qualty due to the elmnaton of muscal and the restoraton of harmonc structure. F. The evaluaton for low bt rate speech codng In ths subsecton, we appled the proposed algorthm whch nvolve wth ptch estmaton and LSFs enhancement n low bt rate speech codng. The speech analyss part s dfferent from the MELP standard. The nosy speech s precleanng based on 465

6 logmmse frstly; the ptch and V/UV decson s confrmed accordng to mproved MBSC algorthm proposed n ths paper. The LSFs s enhanced through DAE model proposed n ths paper before vector quantzaton. The parameter quantzaton and speech synthess s the same as MELP standard. The PESQ scores are shown n Table.4. Table 4 the PESQ results for MELP-400 ose type Method 0dB 5dB 10dB Match Whte Orgnal MELP Precleanng+MELP Proposed Orgnal MELP Precleanng+MELP Proposed Orgnal MELP Precleanng+MELP Proposed The PESQ of match s the average PESQ of four type whch nclude pnk, factory, buccaneer and volvo. It s noted that the PESQ MOS score s hgher wth proposed method n varous condtons. In comparson wth the MELP standard, the proposed method could acheve better objectve speech qualty due to the mprovement of ptch estmaton and the LSFs enhancement. IV. COCLUSIO AD FUTURE WORK In ths paper, we present a speech enhancement approach based on analyss synthess framework. The proposed approach takes advantage of the analyss synthess framework to effectvely elmnate muscal. It bulds the mappng relatonshp to obtan enhanced spectral envelope through deep layer network. The could be suppressed and the dstorted speech could be restored effectvely. The dfferent evaluaton results demonstrate the effectveness of the proposed approach over conventonal approaches n varous nosy condtons. In the future, we wll mprove the current speech enhancement system and focus on adaptaton n real envronment. In addton, we wll consder herarchcal neutral network structure accordng to pror knowledge. We also wll consder mprovng the vocoder structure and expand our algorthm to wdeband speech. ACKOWLEDGMET Ths work s supported by the Major Program for the atonal Socal Scence Fund of Chna (13&ZD189), the atonal atural Scence Foundaton of Chna (SFC) (o , o , o , o , o and o ). REFERECES [1] S. Boll, Suppresson of acoustc n speech usng spectral subtracton, n Acoustcs, Speech and Sgnal Processng, IEEE Transactons on, 1979, 7(): [] Y. Ephram and D. Malah, Speech enhancement usng a mnmummean square error short-tme spectral ampltude estmator, n Acoustcs, Speech and Sgnal Processng, IEEE Transactons on, 1984, 3(6): [3] R. F. Chen, C. F. Chan, and H. C. So, ose suppresson based on an analyss synthess approach, n Proc. Eur. Sgnal Process. Conf. (EUSIPCO), Aug. 010, pp [4] A. Chevegne, H. Kawahara. YI, a fundamental frequency estmator for speech and musc. J. Acoust. Soc. Am. 111, , 00 [5] A. Camacho, J. Harrs. A sawtooth waveform nspred ptch estmator for speech and musc. J. Acoust. Soc. Am. 14, , 008. [6] J. Rouat, Y. Lu, D. Morssette. A ptch determnaton and voced/unvoced decson algorthm for nosy speech. Speech Communcaton. 1, , [7] L.. Tan, A. Alwan. ose-robust F0 estmaton usng SRweghted summary correlograms from mult-band comb flters. In: Proc. IEEE ICASSP, pp , 011 [8] T. V. Sreenvas and P. Krnapure, Codebook constraned wener flterng for speech enhancement, IEEE Trans. Speech Audo Process., vol. 4, no. 5, pp , Sep [9] R. F. Chen, C. F. Chan and H. C. So, Model-Based Speech Enhancement Wth Improved Spectral Envelope Estmaton va Dynamcs Trackng, IEEE Trans. Speech Audo Process., vol. 0, no. 4, pp , Sep. 01 [10] T. Toda, H. Saruwatar, and K. Shkano. Voce converson algorthm based on gaussan mxture model wth dynamc frequency warpng of straght spectrum, In Proc of ICASSP, 001,pp [11] K. Y. Park, H. S. Km. arrowband to wdeband converson of speech usng GMM based transformaton. Proceedng of IEEE Internatonal Conference on Acoustcs, Speech, Sgnal Processng, 000:4: [1] G. E. Hnton, S. Osndero, Y. W. Teh, A fast learnng algorthm for deep belef nets, n eural computaton, 006, 18(7): [13] Y. Bengo, L. Yao, G. Alan, et al, Generalzed denosng autoencoders as generatve models n Advances n eural Informaton Processng Systems, USA, pp: , 013 [14] L.. Tan, A. Alwan. Mult-band summary correlogram-based ptch detecton for nosy speech. In: Speech communcaton, pp , 013. [15] Y. ephram and D. malah, Speech Enhancement Usng a Mnmum Mean-Square Error Log-Spectral Ampltude Estmator, IEEE Trans. Speech Audo Process., vol. 33, no., pp , Sep [16] L. M. Supplee, R. P. Cohn,J. S. Collura and A. V. McCree, MELP: the new federal standard at 400bps In Acoustcs Speech and Sgnal Processng, Germany, ,1997 [17] J. S. Garofolo, TIMIT: Acoustc-phonetc Contnuous Speech Corpus, Lngustc Data Consortum, [18] Rce Unversty, OISEX-9 Database, [Onlne] Avalable: [19] W. Chu, A. Alwan, reducng F0 frame error of F0 trackng algorthm under nosy condton wth an unvoced/voced classfcaton frontend In Acoustcs Speech and Sgnal Processng, Germany, ,009. [0] A. W. Rx, J. G. Beerends, M. P. Holler, et al, Perceptual evaluaton of speech qualty (PESQ)-a new method for speech qualty assessment of telephone networks and codecs, n Acoustcs, Speech, and Sgnal Processng (ICASSP), USA, pp , 001 [1] D. Talkn. Speech Codng and Synthess, Elsever, pp , [] F. Jabloun and B. Champagne. Incorporatng the human hearng propertes n the sgnal subspace approach for speech enhancement. IEEE Trans. on Speech and Audo Processng, 11(6), , [3] B. Kotnk, H. Hoge, and Z. Kacc, "Evaluaton of Ptch Detecton Algorthms n Adverse Condtons". Proc. 3rd Internatonal Conference on Speech Prosody, Dresden, Germany, pp ,