Analysis and Comparison of Evolutionary Algorithms applied to Adaptive Noise Cancellation for Speech Signal

Transcription

1 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) Analyss and Comparson of Evolutonary Algorthms appled to Adaptve Nose Cancellaton for Speech Sgnal Rtka Thakur 1, Papya Dutta 2, Dr. G.C. Manna 3 1 M.Tech (Dgtal Communcaton) Gyan Ganga College of Technology, Jabalpur, M.P, Inda 2 Assocate Professor (Electroncs and Communcaton Dept.) Gyan Ganga College of Tech. Jabalpur 3 Chef General Manager, B.S.N.L, Jabalpur, M.P, Inda Abstract In ths paper, an mproved method based on evolutonary algorthm for speech sgnal denosng s proposed. In ths approach, the stochastc global optmzaton technques such as Artfcal Bee Colony(ABC), Cuckoo Search (CS)algorthm, and Partcle Swarm Optmzaton (PSO) technque are exploted for learnng the parameters of adaptve flterng functon requred for optmum performance. It was found that the ABC algorthm and Cuckoo Search algorthm based speech denosng approach gve better performance n terms of sgnal-to-nose rato (SNR) as compared to PSO based speech denosng approach. The quanttatve (SNR, MSE and Maxmum Error (ME)) and vsual (Denosed speech sgnals) results show superorty of the proposed technque over the conventonal and state -of-art speech sgnal denosng technques. All proposed methods have been smulated n Matlab, and desgn results are llustrated clearly to show the superorty of the proposed method. Keywords Adaptve flters, Adaptve Algorthms, Artfcal Bee Colony, Cuckoo search, Partcle Swarm Optmzaton Algorthm, Speech Enhancement. I. INTRODUCTION The problem of nose cancellaton has receved consderable attenton n recent years n many speech applcatons, such as speech bandwdth compresson, speech recognton, speaker verfcaton, and [1]-[3]. For example, n automatc speech recognton from nosecorrupted speech, the nose cancellaton schemes provde an mproved qualty of speech sgnal that helps n achevng a better recognton performance. The most common problem n speech processng s the effect of nterference nose n the sgnals. Ths nose masks the speech sgnal and reduces ts ntellgblty. Varous sources whch produce nterferences may nclude ventlaton equpment, traffc, crowds and commonly, echoes and reverberaton. It can also arse electroncally from thermal nose, dstorton products or tape hss. If the sound system has large peaks n ts frequency response, the speech sgnal can end up maskng tself [2]. 172 Quanttatve measurements and computer-aded analyss become dffcult and unrelable due to poor speech sgnal qualty. Thus, denosng and enhancement of the speech become necessary for many practcal applcatons. Nose reducton or speech enhancement algorthms are used to suppress background nose and mproves the qualty and ntellgblty of speech. Conventonal lnear flterng s the most common way to do sngle tral analyss of speech sgnal, by whch contamnatons due to on-gong background nose can be attenuated from the speech sgnal. A major dffculty n conventonal lnear flterng s very low SNR, snce then, the concept of adaptve flterng was ntroduced. Adaptve Flters: An adaptve flter adapts tself to changes n ts nput sgnals automatcally accordng to a gven algorthm. The algorthm wll change the flter coeffcents accordng to a gven crtera, typcally an error sgnal s gven to mprove ts performance. In essence an adaptve flter s a dgtal flter combned wth an adaptve algorthm, whch s used to modfy the flter coeffcents and work as an adaptve nose canceller (ANC). Adaptve flters are used n many dverse applcatons lke radar sgnal processng, telephone echo cancellng, equalzaton of communcaton channels and bomedcal sgnal enhancement [4] [5], [6]. Adaptve Algorthms: Adaptve nose cancellaton (ANC) uses varous mnmzaton technques or adaptve algorthms lke LMS, NLMS and RLS. These adaptve algorthms are Gradent based algorthm whch are most commonly used due to smplcty n computaton and easy mplementaton. These gradent based algorthms suffers from some problem lke these are not sutable for multmodal error surface and t gves only one possble soluton for each teraton accordng to generated error. The am of ths paper s to solve the problem of ANC, but these problems are not solved by conventonal algorthm so here optmzaton algorthms (PSO, ABC and cuckoo search) are used. Optmzaton algorthm ncreased the probablty of encounterng the global optmum.

2 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) Performance of the Cuckoo Search based denosng technque performs better than PSO based method for speech denosng. The vsual and quanttatve results are gven n results and dscussons secton. The paper s organzed as follows. Secton-II gves an overvew of ANC for speech denosng. Secton-III ntroduces PSO, ABC and the Cuckoo Search Algorthm for speech denosng case. In secton-iv, qualtatve and quanttatve results of the proposed method wth PSO, supported by SNR, MSE and ME have been dscussed. Conclusons are gven n the fnal secton. II. OVERVIEW OF ADAPTIVE NOISE CANCELLATION Nose cancellaton technology s a growng feld that captalzes on the combnaton of dsparate technologcal advancements. Ths ams to cancel or at least mnmze unwanted sgnal and so to remedy the excess nose that one may experence. There are already several solutons avalable [7], [8]. Adaptve nose cancellaton s wdely used to mprove the Sgnal to Nose Rato (SNR) of a sgnal by removng nose from the receved sgnal. The basc structure and workng of ANC s shown n Fg.1, In ths confguraton the nput x(n) s prmary nput sgnal fed nto transversal flter and generates fltered output y(n) whch s shown as y(n) = w(n) H x(n).the flter output y(n) s compared wth the reference sgnal d(n) and produces error e(n), mathematcal representaton s e(n)=d(n) y(n). Ths error s used to update new flter coeffcents w (n) for transversal flter. Input x(n) TRANSVERSAL FILTER (FIR) w(n) UPDATE ALGORITHM (PSO, ABC, CS) Output y(n) Error e (n) Desred sgnal d(n) Fg. 1 Adaptve Nose Cancellaton Confguraton The error sgnal should converge to the desred sgnal d(n), but t wll not converge to the exact desred sgnal. In other words, the dfference between the desred sgnal d(n) and the error sgnal e(n) wll always be greater than zero. The only opton to ths problem s to mnmze the dfference between those two sgnals usng certan error mnmzaton technques [5]. + III. DIFFERENT EVOLUTIONARY TECHNIQUES FOR DENOISING OF SPEECH SIGNAL Swarm ntellgence (SI) s an nnovatve technque for problem optmzaton whch s nspred by the socal behavours of anmal swarms n nature, examples of systems lke fsh schoolng, brd flockng, ant colones, bactera moldng and anmal herdng. SI s an artfcal ntellgence technque based on anmal behavor n whch no centralzed control (whch shows how ndvdual agents should behave) and self-organzed systems. The Global behavor s acheved by local nteractons between agents. Recent years, several swarm ntellgence algorthms have been proposed [9]. Swarm based algorthm are used n varous applcatons to solve dfferent knd of problems whch has many possble solutons. A. Partcle Swarm Optmzaton Technque Partcle Swarm Optmzaton was proposed n 1995 by James Kennedy and R. C. Eberhart. PSO s an Evolutonary Swarm technque based on stochastc optmzaton and nspred by socal behavor of brd flockng. Here brds fnd the food by the socal cooperaton wth other brds. They move n random drecton to fnd a food and all brd follow the postons and who s closest to food wll nform to other one of the member, but there s no leader. All of them hold ther and all other brds wll update ther poston. Ths process wll contnue untl they reach to food locaton [10-12]. A.1) PSO Algorthm wth Adaptve Flter: The man motve of adaptve flter wth PSO algorthm s to optmze (mnmze) the cost functon whch s calculated by Mean squared dfference.e. MSE, between desred sgnal and fltered output. Ths value s used to calculate ftness value of each partcle. The cost functon or MSE C (n) (the MSE for the th partcle at the nth teraton) s defned as [10, 13-15]: M 1 C n [ e n ] (1) 2 M l 1 l Where, e l (n) s l th error for th partcle and M s number of nput sample. A.2) Procedure of PSO algorthm wth adaptve flters: The PSO algorthm s start wth ntalzaton of partcles and ther velocty wth random swarm wthn a problem space and remanng procedure of PSO algorthm wth ANC can be descrbed n followng steps [16]: 173

3 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) Step1.Start: Decde the problem space and defne the maxmum and mnmum lmt for partcle (±X ).e. weght vector and ther velocty vector (±V ). Step2. Intalze partcles X wth random postons and ther correspondng veloctes V accordng the problem space. Step3. Update Poston: If present poston s wthn the problem space than follow the next step otherwse adjust the postons. Step4.Ftness functon: Calculate ftness functon usng Eq.(1) For each partcle: Calculate P best : Compare ftness value of partcle wth ts P best,, f current value s better then set P best equal to the current value. Calculate G best : Fnd present global mnmum from best poston of partcles, and assgn value to the varable g (G best ). Step5. Update V and X : Update the velocty and poston of the partcle: V pd ( 1) W() V pd () c 1 r 1 (X ppd () X pd ()) c 2 r 2 ( X gpd () X pd ()) (2) X pd ( 1) X pd ( ) V pd ( 1) (3) X represents the th partcle; P represents the best prevous poston; V represents the velocty; c 1 and c 2 are postve constant commonly set to 2. The r 1 and r 2 are two random numbers wthn range 0 to 1. F pp () and F gp () are the personal best poston and global best poston. F pp () s ntalzed wth F pd, whch s calculated n step 4 and the best of F pp () s taken F gp () at ntal step. X ppd () and X gpd () are the personal best poston and the global best poston respectvely. The poston of F pp () s stored n X ppd () and poston of F gp () s stored n X gpd (). Step6. The updatng of poston and velocty s restrcted by the boundary value whch s 80% of maxmum and mnmum value of the partcle search space. Step7. Calculate the ftness F pd () at new search poston. Update F pp () f current value of F pd () s less than the current value of F pp (), otherwse retan old F pp (). Update F gp () f best of F pp () s less than prevous F gp () otherwse retan F gp (), smlarly X ppd () and X gpd () s updated accordngly. 174 Step8. Repeat step 5 to step 7 tll stoppng crtera or maxmum number of teraton s acheved. B. ABC Algorthm Artfcal Bee Colony (ABC) s one of the recently defned algorthms by D. Karaboga n 2005 [17]. The algorthm s specfcally based on the model proposed by Tereshko and Loengarov (2005) for foragng the behavour of honey bee colones. It s motvated by the ntellgent behavour of bees. The ABC algorthm s as smple as PSO and dfferental evolutonary algorthms, and s also easy to mplement. ABC algorthm uses the common control parameters lke colony sze and maxmum number of cycle. ABC as an optmzaton tool provdes a populaton based search n whch ndvduals known as foods postons are modfed by the artfcal bees wth tme. The bee s am s to search or to dscover the places of food sources wth hgh nectar amount and fnally the one wth the hghest nectar. Intally, all food source postons are searched by scout bees and then the nectar of food sources are exploted by the employed bees and onlooker bees, and ths contnual explotaton wll ultmately cause them to become exhausted. In ABC algorthm, the colony of artfcal bees contans three groups of bees: employed bees, onlookers and scouts. The Onlooker bee wats on the dance area for makng decson to choose a food source. A bee gong to the food source vsted by t prevously s named an employed bee. A bee carryng out random search s called as scout. The detaled algorthm steps for the optmzaton of the (thresholdng) functon are as follows: Step1: Intalzaton of control parameters. The man control parameters of the ABC algorthm s colony sze and number of teratons. So, ntalze the colony (CS) sze by 6.Lmt of the scout bees s gven by L= (CS*D)/2. The dmenson D of the problem s three. Step2: Intalze the poston of CS/2 food source of employed bees, randomly usng the varables over a defned range. The range s taken as λ=1 to 150, k=0.1 to 2, m=1 to 4 and fnd CS/2 solutons. Step 3: Evaluate ftness for each of the obtaned soluton. M 1 C n [ e n ] (1) 2 M l 1 l

4 Ftness functon: ftness Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) 1, f C 0 1 C 1 abs C f C 0 (4) Step 4: Select the maxmum value of the ftness, whch s the best qualty of food source. Cycle 1: Employed bees phase v,j = x,j + φ,j (x,j x,j ) (5) Where, k and j are random selecton ndex, ϕ s randomly produced number n the range [0, 1]. Calculate C(v ) and the ftness of v. By comparng wth the ftness of prevously obtaned soluton wth the C(v ). Replace C wth C(v ) f C <C(v ) replace t wth C(v ) otherwse ncrease the tral counter. Step 5: calculate the probablty values p for the solutons x by means of ther ftness usng the equaton: p ftness (6) CS /2 1 ftness Step 6: Onlooker bee s phase: Produce the new soluton v for onlooker bees from the solutons x selected, and dependng upon the probablty p evaluate them. Then apply greedy selecton between C(v ) and C. Step 7: memores the best soluton and correspondng vector. Step 8: scout bee phase, for replacng the abandoned soluton (the soluton of whch tral s more than L), a new random soluton s generated. Step 9: update cycle and repeat the process untl stoppng crtera s acheved. C. Cuckoo Search Algorthm In 2009, Yang and Deb have proposed the Cuckoo Search (CS) optmzaton algorthm [18]. The CS algorthm s a new meta-heurstc algorthm for solvng the optmzaton problems. It s nspred by the oblgate brood parastsm of some Cuckoo speces by layng ther eggs n the nests of other host brds (of other speces). Some host brds can engage drect conflct wth the ntrudng Cuckoos. The algorthm s based on the oblgate brood parastc behavor of some cuckoo speces n combnaton wth the L evy flght behavor of frut fles and some of the brds. 175 Each egg n cuckoo search algorthm represents a soluton and Cuckoo egg represents a new soluton. Overall there s am to employ new and potentally better solutons to replace weak solutons n the nests. In the smplest form, each nest has one egg. The Cuckoo algorthm can be extended to more complcated cases, whch are havng more than one egg representng a set of solutons. The CS s based on three dealzed rules: 1. Each Cuckoo lays one egg at a tme, and dumps t n a randomly chosen nest; 2. The best nests wth hgh qualty of eggs (solutons) wll carry over to the next generatons; 3. The hosts nests are fxed n count, and a host can dscover an alen egg wth probablty p a [0 1]. In ths case, the host brd can ether throw the egg away or abandon the nest to buld a completely new nest n a new locaton. Based on above mentoned rules, the basc steps of Cuckoo algorthm are: Step 1: Set the number of nest. Nest s nothng but dfferent solutons. In ths problem, t s taken as 20. Set the probablty wth a dscovery rate (probablty). Set the stoppng crtera, whch s ether fxed number of teraton or tolerance value. Set dmenson of the problem. The number of dmenson s 3 here. Also set the boundares of the parameters. Step 2: Randomly ntalze the soluton, by generatng n dfferent nest for obtanng n dfferent solutons. Step 3: Evaluate ftness for each of the obtaned soluton. Fnd the best nest correspondng to mnmum value of ftness. Step 4: Start teraton, generate new nest by Levy flght but keep the current best. A Levy flght can be formed as Cuckoo, a Levy flght s performed by the equaton: 1 ' x t x t L evy (7) Where, α s step sze. It essentally provdes a random walk whle the random step length s drawn from L evy dstrbuton, whch has an nfnte varance wth an nfnte mean. L evy dstrbuton s gven by: L evy u = t -λ, (1< λ 3) (8) L evy functon can be changed accordng to applcaton. Mantegna's algorthm s one of the L evy functon.

5 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) Step 5: Evaluate ths set of solutons and obtan the new ftness. Compare old ftness wth ths new ftness, and replace old ftness f new ftness s better than old one. Update the best nest correspondng to ftness. Step 6: Repeat the above process untl some stoppng crtera s acheved gvng the best ftness and correspondng best nest. A complete flowchart routne of optmzaton based methodology for speech denosng s shown n fg.2 depctng the detal steps of overall algorthm. IV. RESULTS AND DISCUSSION In ths secton, varous evolutonary algorthms based on the proposed denosng scheme are exploted for speech sgnals. Performance of the proposed scheme s computed by determnng dfferent fdelty parameters such as SNR, MSE and ME gven by Eqs (9), (10) and (11) respectvely In ths paper, denosng performance of PSO algorthm, ABC algorthm and Cuckoo Search (CS) algorthm s compared on the bass of SNR, MSE and ME. For makng ths comparson, PSO algorthm, ABC algorthm and CS algorthm s executed wth MATLAB R2012a. The performance result of varous evolutonary algorthms s shown n Fg.3, Fg.4, Fg.5 and Table I, Table II and Table III shows the results obtaned by runnng the algorthms for an nput SNR of 5 db, 10 db and 15 db respectvely. It was observed that the ABC algorthm and CS algorthm gves better value of SNR, MSE and ME as compared to PSO. Fdelty parameters: () Desred _ Sgnal 2 SNRdB 10log 10 ( ) (9) Error N 1 MSE Desred _ Sgnal Error (10) () 2 N 1 () ME = max [abs (Desred_ Sgnal Error)] 2 (11) Fg. 2 Flowchart of the optmzaton based methodology for speech denosng. TABLE I For 5 db Input SNR No. of Iteratons PSO Algorthm ABC Algorthm CS Algorthm SNR(dB) MSE ME SNR(dB) MSE ME SNR(dB) MSE ME e e e e e e e e e

6 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) TABLE II For 10 db Input SNR No. of Iteratons PSO Algorthm ABC Algorthm CS Algorthm SNR(dB) MSE ME SNR(dB) MSE ME SNR(dB) MSE ME e e e e e e e e e e e-004 TABLE III For 15 db Input SNR e e e-008 No. of Iteratons PSO Algorthm ABC Algorthm CS Algorthm SNR(dB) MSE ME SNR(dB) MSE ME SNR(dB) MSE ME e e e e e e e e e e e e e e e e-009 Orgnal speech sgnal Denosed by CS algorthm Fg.3 Result of varous evolutonary algorthms for 400 teratons (at 5 db nose) 5 db Nosy sgnal Orgnal speech sgnal Denosed by PSO algorthm 10 db Nosy sgnal Denosed by ABC algorthm Denosed by PSO algorthm 177

7 Internatonal Journal of Recent Development n Engneerng and Technology Webste: (ISSN (Onlne)) Volume 3, Issue 1, July 2014) Denosed by ABC algorthm Denosed by CS algorthm Fg.4 Result of varous evolutonary algorthms for 400 teratons (at 10 db nose) Orgnal speech sgnal 15 db Nosy sgnal Denosed by PSO algorthm Denosed by ABC algorthm Denosed by CS algorthm Fg.5 Result of varous evolutonary algorthms for 400 teratons (at 15 db nose) V. CONCLUSION In ths work, an mproved denosng scheme for speech sgnal s presented usng dfferent evolutonary algorthms such as CS algorthm, ABC algorthm and PSO algorthm. The fdelty parameters obtaned clearly show superorty of the proposed technque over other conventonal speech denosng technques. A comparatve study of dfferent evolutonary algorthms has also been made, and t has been found that the proposed technque based on CS algorthm and ABC algorthm yelds better performance as compared to PSO n terms of SNR, MSE and ME. 178 REFERENCES [1] J.M. Gorrz, J. Ramrez, S. C. Alvarez, C. G. Puntonet, E. W. Lang, and D. Erdogmus, Jan A novel LMS algorthm appled to adaptve nose cancellaton, IEEE Sgnal Processng Letters, vol. 16, pp , [2] Ben G., Nelson M., 2000 Speech and Audo Sgnal Processng John Wley and Sons, Inc. [3] M.S.Ahmed, Jan Comparson of nosy speech enhancement algorthms n terms of LPC perturbaton, IEEE Trans. Acoust., Speech, Sgnal Processng, vol. 37, pp , [4] Smon Haykn, September 2001 Adaptve Flter theory, 4thed. Prentce Hall. [5] Allam Mousa, Marwa Qadus, Shern Bader, Aprl Adaptve Nose Cancellaton Algorthms Senstvty to Parameters, The 2nd Internatonal Conference on Multmeda Computng and Systems (ICMCS'11), IEEE Catalog Number CFP731-3-CDR, ISBN: , pp. 1-5, Ouarzazate, Morocco, [6] Coln H Hansen, Understandng Actve Nose Cancellaton, Taylor and Francs Inc. Prnt ISBN: , [7] V.R.Vjaykumar, P.T.Vanath, P. Kanagasapabathy, Modfed Adaptve Flterng Algorthm for Nose Cancellaton n Speech Sgnals, Electroncs and Electrcal Engneerng, 2007 No 2(74) pp , [8] Xaohu Y., Yunlong Z., Hao Z., Chen H. And Ben N., 2010 An Adaptve Bacteral Foragng Optmzaton Algorthm wth Lfecycle and Socal Learnng, Hndaw Publshng Corporaton Dscrete Dynamcs n Nature and Socety, [9] Mtul K. A., Anl K., Grsh K. S., 2012 Analyss and testng of PSO varants through applcaton n EEG/ERP adaptve flterng approach, Bomedcal Engneerng Letters,, Vol. 2, [10] R. Eberhart, Y. Sh and J. Kennedy, 2001 Swarm Intellgence. San Mateo, CA: Morgan Kaufmann. [11] Dan Palup Rn, St Maryam Shamsuddn and St Sophyat Yuhanz, 2011 Partcle Swarm Optmzaton: Technque, System and Challenges, Internatonal Journal of Computer Applcatons,, Vol. 14(1), [12] D.J. Krusensk and W.K.Jenkns, 2003 Adaptve Flterng Va Partcle Swarm Optmzaton, Conf Sgnal Syst Computer. Vol. 1, [13] Upal Mahbub, Cela Shahnaz and Shakh Anowarul Fattah, 2010 An Adaptve Nose Cancellaton Scheme Usng Partcle Swarm Optmzaton Algorthm, IEEE Internatonal Conference on Communcaton Control and Computng Technologes, Vol.71, [14] Al T.Al- Awam, Azzedne Zergune, Lahouar Cheded, Abdelmalek Zdour, Waleed Saf, 2010 A new modfed partcle swarm optmzaton algorthm for adaptve equalzaton, Dgtal Sgnal Processng, [15] A. Chatterjeea, P.Sarry, 2006 Nonlnear nerta weght varaton for dynamc adaptaton n partcle swarm optmzaton, Computers and Operatons Research, Vol. 33, [16] Karaboga, D.:2005 An Idea Based On Honey Bee Swarm for Numercal Optmzaton, Techncal Report-TR06, Ercyes Unversty, Engneerng Faculty, Computer Engneerng Department. [17] Yang X.S., and Deb, S.:2009 Cuckoo search va Lévy flghts, World Congress on Nature & Bologcally Inspred Computng (NaBIC 2009). IEEE Publcatons. 2009, pp