A Heuristic Speech De-noising with the aid of Dual Tree Complex Wavelet Transform using Teaching-Learning Based Optimization

Transcription

1 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) A Heurstc Speech De-nosng wth the ad of Dual Tree Complex Wavelet Transform usng Teachng-Learnng Based Optmzaton D. Yugandhar #, S.K. Nayak * # E.C.E Department, Adtya Insttute of Technology and Management, Tekkal, Andhra Pradesh, Inda, yugndasar@gmal.com * Electronc Scence Department, Berhampur Unversty, Berhampur, Odsha, Inda sknayakbu@redffmal.com Abstract In our present work, we propose a nature nspred populaton based speech enhancement technque to fnd the dynamc threshold value usng Teachng-Learnng Based Optmzaton (TLBO) algorthm by usng shft nvarant property of dual tree complex wavelet transform (DT-CWT). The performance of these proposed methods are evaluated n terms of Perceptual Evaluaton of Speech Qualty (PESQ) and Peak Sgnal to Nose Rato (PSNR). Speech qualty of dfferent speech waves are compared for two level wavelet packet decomposton and dual tree wavelet transform usng soft threshold. The speech qualtes of the waves are better than the other avalable artcles n the lterature. Keyword- Speech enhancement, Dual tree complex wavelet transform, Teachng-Learnng-based optmzaton, Perceptual Evaluaton of Speech Qualty I. INTRODUCTION In many speech related systems, the orgnal speech sgnal s contamnated wth some nterference sources. The nterference source,.e the nose sgnal degrades the qualty of the orgnal speech. The speech sgnal s affected by wde-band nose n the form of whte or colored nose and a perod nose such as hum nose. The most common type of nose n communcaton channels s the addtve wde band Gaussan nose. Speech enhancement ams at mprovng the performance of speech communcaton systems n nosy envronments. Speech enhancement may be appled, for example, to a moble rado communcaton system, a speech recognton system, a set of low qualty recordngs, or to mprove the performance of ads for the hearng mpared. Several methods []-[6] have been proposed n the lterature for the enhancement of degraded speech. A maorty of these methods can be grouped nto spectral processng and temporal processng methods. In the spectral processng methods, degraded speech s processed n the frequency doman mostly usng Fourer transform for achevng speech enhancement. In the temporal processng methods, the processng s done n the tme doman. Most speech enhancement methods mprove the qualty of the sgnal but degrade ts ntellgblty of the speech. Performance measures lke PSNR and PESQ are wdely used as the performance of the evoluton crteron. For elmnaton of the Gaussan back ground nose n the communcaton channels, we have been mplementng adaptve thresholdng technque usng TLBO optmzaton. The TLBO algorthm [7]-[9] s a global optmzaton, populaton based teratve learnng mechansm that exhbts some common characterstcs wth other evolutonary computaton (EC) technques lke (GA) [], Partcle Swarm Optmzaton (PSO) [], Dfferental Evoluton (DE) [], and Artfcal Bee Colony (ABC) [3]. The TLBO algorthm does not requre any algorthm-specfc control parameters lke mutaton and crossover as n genetc algorthm. The TLBO methods provde the learnng mechansm n adaptve models. The organzaton of ths paper s as follows. Secton II presents speech de-nosng usng wavelet thresholdng. Secton III dscusses ntroducton of TLBO algorthm, Secton IV dscusses the proposed method of mplementaton. The smulaton results and comparsons are gven n secton V. Fnally the paper concludes wth secton VI. II. SPEECH DE-NOISING USING WAVELET THRESHOLDING The Fourer transform s the basc tool n frequency doman approach. Short tme Fourer transform (STFT) [4] provdes one of many ways to generate a tme frequency analyss of sgnals. However, fxed tme-frequency resoluton of the STFT poses a serous constrant because the wdth of the wndow functon s fxed. A varable wndow sze s requred to represent sgnals at dfferent resoluton. Another lnear transform that provdes such a tme frequency analyss s known as contnuous wavelet transform (CWT). Ths CWT gves a wndow functon whose wdth ncrease n tme whle resolvng the low-frequency contents, and decreases n tme whle resolvng DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 967

2 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) the hgh-frequency contents of a sgnal, whch leads to good tme-frequency resoluton. The CWT of a functon f ( t) L wth respect to some analyzng wavelet s gven by [5] W ( b, a f b, a) f ( t) ( t) dt t b where, b, a ( t) ( ), a () a a The parameters b and a n () are the translaton and dlaton (scalng) parameters respectvely. The normalzaton factor s ncluded so that b, a. The expresson of the nverse transform s to a reconstruct the orgnal functon from ts ntegral wavelet transform s gven by f ( t) db [ W f ( b, a)] b, a( t) da C (3) a where, C s a constant that depends on the selecton of the wavelet and s gven by ˆ ( ) C d The condton above s known as the admssblty condton, restrcts the class of functons lke wavelets. Fndng nverse wavelet transformaton for syntheszng the orgnal sgnal n (3) s qute cumbersome because t nvolves a two dmensonal ntegraton over the scale parameter a and the translaton parameter b. If we consder scale parameter a to be of the form s and translaton parameter b of the form k s, where k, s, wth these values of a andb, the dscrete form of ntegral n () representaton s known as dscrete wavelet transform (DWT). s s s / s W f ( k, ) f ( n) ( n k) (5) n Let us consder the dscrete functon f m (n) whch s delayed by m samples represented as f m ( n) f ( n m), and then ts DWT s computed as gven n (6). The mportant observaton from (5)-(6) s that the DWT of a functon shfted n tme s qute dfferent from the DWT of the orgnal functon. Therefore, the DWT transform s shft varant transformaton. The wavelet transform can be regarded as a bank of bandpass flters wth constant Q factor. The advantage of usng varable sze wndows for dfferent frequency bands s seen n the wavelet analyss. s s s / s W f ( k, ) f ( n m) ( n k) m W s / n n s s f ( n) [( n ( k m )] f [( k m s ) The wavelet analyss has a dstnct ablty to detect local features of the sgnal n both tme and frequency, such as the plosve fne structures of the speech and other transent, nstantaneous and dynamc speech components that contrbute sgnfcantly to the qualty of the speech. Therefore, wavelet transform can provde an approprate model of speech sgnal for de-nosng applcatons. One popular technque for wavelet based sgnal enhancement s the wavelet shrnkage algorthm [6]. Wavelet shrnkage s a smple de-nosng method based on the thresholdng of the wavelet coeffcents. The estmate threshold defnes the lmt between the wavelet coeffcents of the nose and those of the target sgnal. However, t s not always possble to separate the components correspondng to the target sgnal from those of nose by smple thresholdng. Applyng thresholdng unformly to all wavelet coeffcents not only suppress addtonal nose but also some speech components lke unvoced ones whch leads to the loss of perceptual qualty of the fltered speech. In order to mprove perceptual speech qualty varous thresholdng and estmaton technques [7] have been proposed. s, s ] () (4) (6) DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 968

3 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) Assumng that s (n) represents the clean speech wth fnte duraton, x (n) stands for the speech corrupted by whte Gaussan nose n (n) s havng zero mean and varance s shown n (7) and correspondng wavelet doman representatons s shown n (8) x( n) s( n) n( n) (7) S N (8) If W denotes a wavelet transform matrx, (8) represents wavelet coeffcents, where W, S Ws, and N W n are nosy speech, orgnal speech and nose wavelet coeffcents respectvely. The de-nosed sgnal Ŝ s obtaned by soft thresholdng nosy wavelet coeffcents whch can be represented as Sˆ THR(, T ) (9) where THR(.) denotes a thresholdng functon and T denotes the threshold value. Thresholdng of coeffcents can be n done n many ways. However there are two popular versons known as hard thresholdng and soft thresholdng. Hard thresholdng sets any wavelet coeffcent whose absolute value s less than or equal to threshold s to zero whle the others are kept unchanged., f T ˆ (), f T The soft thresholdng s smlar to the hard thresholdng except that t ether shrnks or klls (set to zero) coeffcents based on the threshold condton gven below. sgn( )( T ), f T ˆ (), f T The Donoho [8] showed that when nose domnates the observed data, the unversal threshold method performs well and when the underlyng sgnal domnates the observed data, the SURE method [9] performs better than unversal Threshold method. Donoho and Johanstone [6] proposed the unversal threshold value as. MAD denotes the absolute medan value on the log( N) and standard devaton MAD/ frst scale of the detaled wavelet coeffcents and N s the length of the nosy sgnal. Shma [], obtaned threshold value basng upon the symmetrc Kullback- Lebler dvergence between the probablty dstrbutons of nose wavelet coeffcents and nosy speech and fnally obtaned the mproved threshold value usng the segment Sgnal to Nose Rato. It can be observed that the soft thresholdng method removes more nose components than the hard thresholdng method whch leads the sgnal degradaton to hgher range. So, an approprate thresholdng technque s to be selected for optmal enhancement of speech sgnals. Besdes ths, some other thresholdng technques lke Sten s unbased estmate selecton rule, Heurstc threshold selecton rule and Mnmax performance threshold selecton rule has been used n normal practce. A. The Dual Tree Complex Wavelet Transform (DT-CWT) The wavelet transform suffers from four fundamental shortcomngs namely oscllatons, shft varance, alasng and lack of drectonalty []. The DT-CWT s a recent enhancement to the DWT wth complex valued scalng functon and complex valued wavelet functon whch can be represented as c c ( t) ( t) ( t) () c r ( t) ( t) ( t) (3) c r where, r and r are real and even functon of t and and are magnary and odd functon of t. However, r and form a Hlbert transform par for wavelet functon and smlarly r and form Hlbert transform for scalng functon. Hence, c and c are both analytc sgnal and supported on only one half of the frequency axs. The DT-CWT ntroduced by Kngsbury, t conssts of two real dscrete wavelet transforms. The frst DWT gves the real part of the transform whle the second DWT gves the magnary part. The two real DWTs use two dfferent sets of perfect reconstructon (PR) flter whch are ontly desgned so that the overall transform s approxmately analytc. The analyss flter banks used to mplement DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 969

4 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) DT-CWT of a sgnal x (n) s shown n Fg. []. The two real wavelet transforms use two dfferent sets of flters, wth each satsfyng the PR condtons. The two sets of flters are ontly desgned so that the overall transform s approxmately analytc. The flters h n), h ( ) denote the low-pass and hgh-pass conugate ( n quadrature flter (CQF) par for the upper bank. The autocorrelaton of the flter can be expressed as n for k h ( n) h ( n k) ( k) (4) for k (n ) h ( n) ( ) h ( n ) Smlarly, the flters g n), g ( ) form another CQF par whch denotes the low-pass and hgh pass flters ( n respectvely for the lower flter bank. The DT-CWT of an nput real vector can be represented by the rectangular matrx gven below F F h g (5) F (6) Let w h Fh and w F g g represents real part and magnary part of the DT-CWT then wh w g represents complex dual tree wavelet coeffcents. Level Level h (n ) h (n h (n ) h (n Real x(n) g (n g (n g (n Imagnary g (n Fg. Flter bank of Two Level Dual Tree Complex Wavelet Transform III. INTRODUCTION OF TLBO ALGORITHM In ths paper, we have adapted TLBO algorthm shown n Fg. [7] to fnd the better threshold value n dfferent scales of DT-CWT. The algorthm s based upon two phases lke teacher phase and learner phase. A. Teacher phase In the teacher phase of the algorthm the knowledge flow s from the teacher to the learners. Durng the teacher phase a teacher tres to ncrease the mean result of the entre class n the subect taught by hm or her dependng on hs or her sklls and knowledge. In the begnnng of the algorthm let say M be the mean of each subect and T be the teacher.e most learned person at any teraton. Now teacher T wll mprove exstng mean M based on hs sklls and knowledge so that the new mean wll be desgnated as M new. The dfference between the new mean and exstng mean s gven as Dfference _ Mean r ( M T M ) (7) new F where r s the random number whch takes the values n the range and, T F known as teachng factor, t s not a parameter of TLBO algorthm. The value of T F s not specfed n the begnnng of the executon of the DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 97

5 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) algorthm. The value of T F s randomly selected by the algorthm tself wth equal probablty usng the equaton (8). Evaluate the ntal populaton Calculate mean of each desgn varable Select the best soluton (Teacher) Modfy soluton based on best soluton Dfference _ Mean r ( M T M new F ) Reect No Is the new soluton s better than exstng? Yes Accept Keep the Prevous soluton Select two solutons randomly and No Is Yes r ( new, old, ) r ( new, old, ) Reect No Is new soluton better than exstng? Yes Accept Keep the Prevous soluton Is termnaton Crteron fulflled? No Yes Fnal value of soluton Fg.. Flowchart of TLBO algorthm. DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 97

6 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) T F round [ rand(,)] (8) From (8), t s concluded that the teachng factor value ( T F ) s n between and. Based on the Dfference _ Mean, the exstng soluton s updated n the teacher phase usng the equaton gven below new, old, Dfferenc_ Mean (9) new, s the updated soluton value of exstng soluton old where,. Accept new, f t gves better ftness value. At the end of the teacher phase all the accepted functon values are stored and these values become the nput to the learner phase. The learner nfluence and knowledge depends on the outcome of the teacher phase. B. Learner Phase It s the second part of the TLBO algorthm where Learners ncrease ther knowledge through nteracton among themselves. A learner nteracts randomly wth other learners to mprove hs or her knowledge. A learner can enhance ther knowledge and able to learn new thngs f the other learner has more knowledge than hm or her. Mathematcally, the learnng phenomenon of ths learnng phase s expressed below. At any teraton, consder two dfferent learners and where r ( ), f F ) F( ) () new, old, ( r ( ), f F ) F( ) () new, old, ( where, F( ), F( ) are known as the learners knowledge or ftness value of the learners and respectvely. IV. PROPOSED METHOD OF IMPLEMENTATION Dfferent speech de-nosng smulatons have been carred out to verfy the de-nosng performance of the speech sgnal usng TLBO algorthm. In these smulatons, orgnal speech sgnals from NOIZEUS speech database were consdered. The nput PSNR of 8.3 db and db are added to the orgnal speech sgnals to conduct two dfferent smulatons. The tranng process to acheve the optmzed threshold value s shown n Fg. 3. The two threshold values obtaned from the tranng process s used as the nput to de-nose speech sgnals corrupted wth nput PSNR of 8.3 db and db s shown n Fg. 4. A. Implementaton of TLBO: Tranng: Step : Clean speeches from sp_.wav to sp_.wav were taken from the database. Step : Nosy nput sgnals are generated by addng nput PSNR of 8.39dB and db. Step 3: Populatons taken as 3 learners and two subects are consdered as desgn varables and teratons were consdered to tran wth TLBO algorthm. Step 4: Each nosy speech sgnal s decomposed usng two level DT-CWT whose analyss and synthess flter coeffcents have shown n Table I & II respectvely. The nverse DT-CWT can be found after soft thresholdng detals wavelet coeffcents by usng (). Step 5: PESQ s selected as ftness functon and the value s measured between correspondng de-nosed and orgnal sgnal of all the ten speech sgnals. The mean PESQ value of all ten sgnals s consdered as ftness functon to be maxmzed. Step 6: Iteratons were conducted to get optmzed threshold. DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 97

7 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) TABLE I Analyss dual tree complex wavelet coeffcents Frst Stage Real Coeffcents Frst Stage Complex Coeffcents Second Stage Real Coeffcents Second Stage Complex Coeffcents LPF HPF LPF HPF LPF HPF LPF HPF Frst Stage Real Coeffcents TABLE II Synthess dual tree complex wavelet coeffcents Frst Stage Complex Coeffcents Second Stage Real Coeffcents Second Stage Complex Coeffcents LPF HPF LPF HPF LPF HPF LPF HPF DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 973

8 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) Tranng phase NOIZUS Tranng Speech Database Nosy Speech Sgnal Nose Dual Tree Complex Wavelet Transform Soft Threshold Iteratve TLBO Algorthm for Threshold Selecton Inverse Dual Tree Complex Wavelet Transform Optmal Threshold Value Fg 3: Archtecture of proposed speech de-nosng system for tranng B. De-nosng of speech sgnals: Testng Step : Nosy speech sgnals sp_.wav to sp_.wav are generated by consderng nput PSNR of 8.39dB and db. Step : The obtaned optmum threshold values from the tranng process are used to de-nose the nosy speech sgnals usng two level DT-CWT by soft thresholdng detals wavelet coeffcents by usng (). The archtecture of the proposed system for de-nosng of speech sgnals wth voce actvty detecton (VAD) has shown n Fg. 4. The VAD method proposed n [][3] to dentfy both voce and slence zones has been mplemented usng Texas nstruments TMS3C673 floatng pont dgtal sgnal processor operatng at 5MHz to check the performance of speech de-nosng n real tme. Step 3: PSNR and PESQ performance measures were calculated between orgnal and de-nosed speech sgnals. C. Performance metrcs of speech de-nosng: Many speech qualty measures were proposed n the lterature [4]. They are segmental Sgnal to Nose Rato (SNR), Itakura-Sato dstance measure, Cepstrum dstance measures, output PSNR and PESQ. ) PESQ: Among all other obectve measures mentoned above, the computatonal burden s more when PESQ measure s computed between orgnal and de-nosed sgnal. Ths measure s recommended by nternatonal telecommuncaton unon standardzaton sector (ITU-T) for speech qualty assessment of 3. khz handset telephony as descrbed [4], the PESQ s computed as a lnear combnaton of the average asymmetrcal dsturbance values A and the average dsturbance value D ) whch can be expressed as nd PESQ a adnd a A nd () where, a 4. 5, a. and a. 39 are constants. The parameters n the above equaton were optmzed for speech processed through networks. ) Output PSNR: Here, we have been evaluatng the performance of the algorthm by consderng output PSNR. It s the rato between the power of output sgnal and the power of nose n decbel scale. PSNR s most commonly used to measure the qualty of reconstructed sgnal. Output PSNR n db *log ˆ( S( n) S n M Sˆ, where ) e e ( nd (3) DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 974

9 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) where M S( n), Sˆ( n),, and, Sˆ sgnal, power of de-nosed sgnal and nose power respectvely. e are the length of the sgnal, orgnal speech sgnal, de-nosed speech Testng Phase Nosy Speech Database Dual tree Complex Wavelet Transform Soft Thresholdng Optmum Threshold Value Obtaned from tranng Inverse Dual tree Complex Wavelet Transform Voce Actvty Detecton De-nosed Speech Sgnal Fg 4: Archtecture of proposed system for speech de-nosng wth VAD V. SIMULATION RESULTS AND COMPARISONS For smulaton purpose we have been consderng a populaton sze of 3 learners, soluton length of and teratons. An extensve evaluatons and comparsons have done by takng two, three and four levels of decomposton of a sgnal n the dual tree complex wavelet doman. Two level DT-CWT gves good result comparatvely other levels of decompostons. The PSNR and PESQ measures were gven for nput PSNR of 8.39dB and db shown n the Table III and Table IV and the correspondng speech sgnal outputs were shown n Fg. 5 & Fg. 6 respectvely. The de-nosed speech sgnal usng TLBO preserves both voced and unvoced speech comparatvely usng conventonal DT-CWT were shown n Fg. 5(d) and 6(d) for nput PSNR of 8.39dB and db respectvely. The output of de-nosed speech sgnals after makng slence zones to zero are shown n Fg. 5(f) and 6(f). The output of VAD usng TMS3C673 s also shown n Fg. 7. Comparsons charts of PSNR and PESQ when nput nose SNR of db was shown n Fg. 8 & Fg. 9. TABLE III Output PSNR and PESQ of all three methods when nput PSNR s 8.39 db S.No Speech Fle Name Wavelet Packet Dual Tree Wavelet TLBO PSNR PESQ PSNR PESQ PSNR PESQ Sp_.wav Sp_.wav Sp_3.wav Sp_4.wav Sp_5.wav Sp_6.wav Sp_7.wav Sp_8.wav Sp_9.wav Sp_.wav DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 975

10 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) S.No Speech Fle Name TABLE IV Output PSNR and PESQ of all three methods when nput PSNR s db Wavelet Packet Dual Tree Wavelet TLBO PSNR PESQ PSNR PESQ PSNR PESQ Sp_.wav Sp_.wav Sp_3.wav Sp_4.wav Sp_5.wav Sp_6.wav Sp_7.wav Sp_8.wav Sp_9.wav Sp_.wav (a) x (b) x 4 Ampltude (c) x (d) x (e) x (f) Sample Number Fg. 5: TLBO wth DT-CWT output for speech corrupted wth whte nose wth nput PSNR 8.39 db: (a) Orgnal speech sgnal (b) Nosy speech sgnal (c) De-nosed speech usng DT-CWT. (d) De-nosed speech usng TLBO wth DT-CWT (e) Voced and slence zones of TLBO output (f) TLBO voce actvty detecton output makng slence zones zero x 4 DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 976

11 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) (a) x (b) x 4 Ampltude (c) x (d) x (e) x (f) Sample Number Fg. 6: TLBO wth DT-CWT output for speech corrupted wth whte nose wth nput PSNR db: (a) Orgnal speech sgnal (b) Nosy speech sgnal (c) De-nosed speech usng DT-CWT. (d) De-nosed speech usng TLBO wth DT-CWT (e) Voced and slence zones of TLBO output (f) TLBO voce actvty detecton output makng slence zones zero. x 4 DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 977

12 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) Fgure 7: VAD output usng TMS3C673 dgtal sgnal processor Fgure 8: Comparson graph of PSNR for nput PSNR db DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 978

13 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) Fgure 9: Comparson graph of PESQ for nput PSNR db VI. CONCLUSION TLBO algorthm s one of the meta- heurstc algorthm has been proposed recently for solvng engneerng optmzaton problems. The soft thresholdng shrnks wavelet coeffcents based on the threshold value leads to decrease n the sgnal energy as well as nose. Therefore, the output PSNR s less than the nput PSNR n the case of wavelet packets and complex dual tree wavelet transform. The threshold obtaned by TLBO algorthm preserves the sgnal energy as well as the qualty of the speech sgnal leads to ncrease n output PSNR as compared to other two methods. TLBO method of de-nosng s able to provde 5%, 8% mprovement n average output PSNR and average PESQ than wavelet packet transform and 3%, % mprovement n average output PSNR and average PESQ than conventonal dual tree complex wavelet transform. The Lstenng tests usng DSK 673 processor show that the de-nosed speech sgnal obtaned by TLBO preservng both voced and unvoced sounds. ACKNOWLEDGMENT Ths research paper s made possble through the help and support from our beloved drector Prof. V.V.N Rao, Prncpal Dr. K.B.Madhu Sahu and TEQIP co-ordnator Dr. D.Vshnu Murty for provdng TMS3C673 dgtal sgnal processor kt wth Code Composer Studo and MATLAB software n the laboratory. I would lke to thank Prof. Ravpud Venkata Rao, Department of Mechancal Engneerng, S.V.N.I.T, Surat, Guarat, Inda for conductng workshop on soft computng technques wth % commtment to hone sklls of young researchers. REFERENCES [] S. Boll, Suppresson of acoustc nose n speech usng spectral subtracton, IEEE Trans. Acoust., Speech, Sgnal processng., vol. 7, no., pp. 3-, 979. [] P.C. Lozou, Speech Enhancement: Theory and Practce, CRC Press, Frst edton, Boca Raton, FL,7. [3] M. Berout, R.Schwartz and J.Makhoul, Enhancement of speech corrupted by acoustc nose, IEEE Proceedngs of Internatonal Conference on Acoust., Speech, Sgnal process., 979, vol. 4, pp. 8-. [4] P.Lockwood, and J. Boudy, Experments wth a nonlnear spectral subtractor (NSS), Hdden Markov Models and the proecton, for robust speech recognton n cars, Speech Communcaton, vol., no. 3, pp. 5-8, 99. [5] S. Kamath and P.A. Lozou, A mult band spectral subtracton methods for enhancng speech corrupted by colored nose, IEEE Proceedngs of Internatonal Conference Acoust., Speech, Sgnal process., vol. 4, pp. IV-464,. [6] Y. Ephram and D. Malah, Speech Enhancement usng a mnmum mean square error short-tme spectral ampltude estmator, IEEE Trans. Acoust., Speech, Sgnal process., vol. 33, no., pp , 985 [7] R.V.Rao, V.J.Savsan and D.P.Vakhara, Teachng learnng-based optmzaton: a novel method for constraned mechancal desgn optmzaton problems, Computer Aded Desgn, vol. 43, no. 3, pp ,. [8] R.V.Rao, V.J.Savsan and D.P.Vakhara, Teachng learnng-based optmzaton: an Optmzaton method for contnuous non-lnear large scale problems, Informaton Scences, vol. 83, no., pp. -5,. [9] R.V.Rao and V. Patel, An eltst teachng learnng-based optmzaton algorthm for solvng complex constraned optmzaton problems, Internatonal Journal of Industral Engneerng Computatons, vol. 3, no. 4, pp ,. [] K.F. Man, K.S.Tang and S.Kwong, Genetc algorthms: concepts and applcaton [n engneerng desgn], IEEE Trans. Industral Electroncs, vol. 43, no. 5, pp , 996. [] J. Kennedy and R. Eberhart, Partcle Swarm Optmzaton, IEEE Internatonal Conference on Neural Network., vol. 4, pp , 995. [] S. Das, P.N. Suganthan, Dfferental Evoluton: A survey of the State-of-the-Art, IEEE Transactons on Evolutonary Computaton, vol. 5, no., pp. 4-3,. [3] F.S. Abu-Mout and M.E. EI-Hawary, Overvew of Artfcal Bee Colony (ABC) algorthm and ts applcatons, IEEE Internatonal System Conference, pp. -6,. DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 979

14 ISSN (Prnt) : ISSN (Onlne) : D. Yugandhar et al. / Internatonal Journal of Engneerng and Technology (IJET) [4] I.Y. Soon and S.N. Koh, Speech enhancement usng a -D Fourer Transform, IEEE Trans. on Speech and audo process., vol., no. 6, pp , 3. [5] J.C. Goswam, and A.K. Chan, Fundamentals of wavelets theory, algorthms & applcatons, John wley & sons, Inc., Publcaton,. [6] D.L. Donoho and I.M. Johnstone, Adaptng to unknown smoothness va wavelet shrnkage, Amercan Statstcal Assoc., vol. 9, no. 43, pp. -4, 995. [7] J.H. Chang, S. Gazor, N.S. Km and S.K. Mtra, Multple statstcal models for soft decson n nosy speech enhancement, Pattern Recognton., vol. 4, no. 3, pp. 3 34, 7. [8] D. Donoho, De-nosng by soft-thresholdng, IEEE Trans. Informaton Theory, vol. 4, no. 3, pp , 995. [9] Y. Hu and P.C. Lozou, Speech enhancement based on wavelet thresholdng the mult-taper spectrum, IEEE Trans. Speech Audo Process., vol., no., pp , 4. [] S. Tabban, A. Akbar and B. Nasersharf, Speech enhancement usng a wavelet thresholdng method based on symmetrc Kullback- Lebler dvergence, Sgnal process., vol. 6, pp , 5. [] D. Herc, and B. Potocnk, Image enhancement by usng drectonal wavelet transformaton, Journal of Computng and Informaton Technology, vol. 4, no. 4, pp , 6. [] Y. Ghanbar, M.R. Karam-Mollae, A new approach for speech enhancement based on the adaptve thresholdng of the wavelet packets, Speech Commun., vol. 48, No. 8 pp , 6. [3] J. Ramírez, J.C. Segura, C. Benítez, A. Torre, A.J. Rubo, A new Kullback Lebler VAD for speech recognton n nose, IEEE Sgnal Process. Lett., Vol., No. ), pp , 4. [4] Y. Hu and P.C. Lozou, Evaluaton of Obectve Qualty Measures for Speech Enhancement, IEEE Transactons on Audo, Speech and language processng, vol. 6, no., pp. 9-38, 8. AUTHOR PROFILE D.Yugandhar completed hs M. Tech. from JNTU, Hyderabad n 7 and regstered for Ph. D. n Berhampur Unversty. He has years of experence n the ndustry and 3 years of teachng experence n varous engneerng colleges. Presently he s workng as Assocate Professor n the Dept. of ECE of Adtya Insttute of Technology and Management, Tekkal, Andhra Pradesh, Inda. Hs nterestng areas of research are Sgnal processng and Image Processng. Dr. S. K. Nayak completed hs M. Tech. from IISC, Bangalore n 99 and Ph. D. degree from Berhampur Unversty n 996. He s now workng as Professor n the Dept. of Electronc Scence, Berhampur Unversty, Berhampur. Hs nterestng areas of research are embedded systems, Image Processng and Sgnal processng. DOI:.87/et/6/v85/68545 Vol 8 No 5 Oct-Nov 6 98