Fuzzy ARTMAP Technique for Speech Noise Reduction

Transcription

1 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, 7 Fuzzy ARTMAP Technique for Speech Noise Reduction BOONRUANG MARUNGSRI nd ANANT OONSIVILAI School of Electricl Engineering, Institute of Engineering Surnree University of Technology University Ave., Mung District, Nkhon Rtchsim, THAILAND mshvee@sut.c.th, nnt@sut.c.th Astrct: - This pper presents n pproch to reduction noise of speech voice commnded utomtic heel chir using technique sed on the fuzzy ARTMAP neurl netork (FAMNN). The mesurle output noisy speech ith db, db, -db, -db nd -db SNR level is otined s the contminted signl of the interference to compre ith the output dt of the filter. The hite noise source is cquired s the input. Finlly, fter trining, the fuzzy ARTMAP output (i.e. estimted interference) s demonstrted. Then the estimted informtion signl is clculted s the difference eteen the mesured signl nd the estimted interference. The fuzzy ARTMAP could do prcticlly superior sitution in dptive de-noising of speech voice commnded utomtic heel chir system ith nonliner chrcteristics. Key-Words: - speech, noise, noise reduction, fuzzy ARTMAP neurl netork, FAMNN Introduction Recently, technologies hve een continuously developed in order to mde comfortle dily life. Technologies not only developed for norml peoples ut lso developed for hndicpped peoples too. Hndicpped peoples hve right to tke cre. By this reson, e hve developed speech voice commnded utomtic heel chir for the hndicpped people s shon in Fig.. Commnd voices re go hed, go ck, turn left, turn right nd stop in Thi lnguge. Side vie Fig. chir. Front vie Speech voice commnded utomtic heel This utomtic heel chir received commnd voice from the commnder vi microphone. The receiving commnd voice s used to compre ith the reference commnd voices. If the receiving voice is sme or ner the sme ith the reference voice then the utomtic heel chir done y using the receiving commnd voice. Since e live in nturl environment here noise is inevitle nd uiquitous, speech signls cn seldom e recorded in pure form nd re generlly contminted y coustic ckground noise. As result, the microphone signls hve to e clened up ith digitl signl processing tools efore they re stored, trnsmitted, or plyed out. The clening process, hich is often referred to s either noise reduction or speech enhncement, cn e chieved in mny different ys. In recent yers, much reserch hs een undertken into noise reduction to improve the performnce of speech communiction nd recognition systems[ ]. Noise reduction suppress environmentl noises, improving the speech qulity nd incresing the recognition ccurcy. Thus, noise reduction hs een of incresed interests for mny reserchers. Fuzzy ARTMAP Neurl Netork The Fuzzy ARTMAP neurl netork (FAMNN) hs een introduced nd detiled description extensively y Crpenter et l. in 99 []. The rchitecture of FAMNN sed on Adptive Resonnce Theory (ART). The fmily of ART NN [] includes the fuzzy ART mpping (ARTMAP) NN (FAMNN), the Fuzzy min mx NN, the lterlly primed dptive resonnce theory (LAPART), the evidence integrtion for dynmic predictive mpping (ART-EMAP), nd the Gussin ARTMAP. The FAMNN hve een successfully pplied in mny tsk such s dt mining [], remote sensing [], recognition of hnd ritten

2 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, 7 chrcters [], nd signture verifiction [6] etc. For completeness in the folloing sections, e present only the necessry detils in this section. One dvntge is tht FAMNN is fster to trin thn other neurl netorks due to the smll numer of trining epochs required y the netork to lern the input dt. FAMNN is considered fst even mong memers of the ARTMAP fmily due to the computtionlly chep mpping eteen inputs nd outputs. Also, the clssifiction results of FAMNN re esily interpretle. Furthermore, compred to stndrd nerest neighor techniques hich re lso commonly used, FAMNN requires less memory since it uses compressed representtion of the dt, nd for the sme reson FAMNN requires less clssifiction time. The existence of memory reduction lgorithms does not diminish this dvntge of the FAMNN since these lgorithms cn e used for oth nerest neighor techniques nd the FAMNN. The FAMNN is supervised netork hich is composed of to Fuzzy ART modules, ART nd ART, s shon in Fig.. Those modules re linked together vi n inter-art module, F, clled mp field. The mp field is used to relize the mtch trcking rule, herey the vigilnce prmeter of ART in crese in response to predictive mismtch t ART. Mtch trcking reorgnizes ctegory structure so tht predictive error is not repeted on susequent presenttions of the input. ART F F F y x A=(, c ) Input, I reset ρ mp field F X ρ Mtch trcking F F F k y x B=(, c ) ART reset Desired Output, I (during trining) Fig. FAMNN processor consists of to fuzzy ART modules linked y mp field. As shon in Fig., vriles in ART nd ART module re designed y suscripts or superscripts nd. Inputs to ech module re in the complement c code form, I = A = (, ) for ART module nd c I = B= (, ) for ART module, respectively. For ART module, x ( x,..., x ) represent the F output vector, y y y (,..., M ) M represent the ρ F output vector nd (,..., M ) represent the th eight vector. For ART module, x ( x,..., x ) represent the F output vector, y y y (,..., M ) M represent the F output vector nd k ( k, k..., k M ) represent the k th eight vector. For the mp field, let x ( x,..., x ) represent the F output vector nd N (,..., N ) represent the eight vector from the th F node to F. Vector x, y, x, y nd x re set into eteen input presenttion. The mp field F is ctivted henever one of the ART or ART ctegories is ctive. If node J of F is chosen, then its eights K of ctivte F is ctive, then the node K in ctivted y pthys eteen y x = y F. If node F nd th If the J F node is ctive nd F is ctive th If the J F node is inctive nd F is ctive If the F is inctive nd F is ctive If the F is inctive nd F is inctive F is F. (.) Due to composed of to Fuzzy ART module, Fuzzy ARTMAP module vectors [7] sed on to seprtely distnce criteri, mtch nd choice. The mtch function is defined y S ( I ) I I (.) here is n nlog-vlued eight vector ssocited ith cluster, denotes the fuzzy AND opertor, (p q) i min(p i,q i), nd the norm... is defined y p =. The choice function is defined y i p i I T ( I ) (.) α + here α is smll constnt. Incresing α ises the serch more tords modules ith lrge. Input vector (I) is ssigned to the ctegory ith mximizes T (I) hile stisfying S (I) ρ, here the vigilnce, ρ, is constnt, ρ. The fuzzy ARTMAP lerning rule is given y old ; ne Ji Ji Ii Ji = (.) old old Ji - β(ji - I i ) ; Ji > Ii here < β. Only the eights of the Cluster to hich I hs een ssigned re updted. All i re initilly set to.

3 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, 7 The mp field is fundmentlly look-up tle, retrieving n nlog-vlued eight hen module node J nd module node L re ctive. Note tht only one node of ech module is ctive t given time. If r, is rised until node J ecomes inctive (nd some other node ecomes ctive). This process is repeted until ρ < ρ the vigilnce of module,. When the next input is presented, r is returned to its seline vlue. All re initilly set to. During lerning, hen nodes J nd L ecome ctive nd ρ, ll, l L, re reduced in vlue (typiclly set to ). Principles of Speech Noise Reduction In generl, the noisy speech signl is given y x( k) = s( k) + n( k ), here s( k ) nd n( k ) represent the clen speech nd noise signl, respectively. Adptive Noise Cncelltion is used to remove noise from useful signls. This is useful technique here signl is sumerged in noisy environment. The process of using the FAMNN for solving the dptive noise cncelltion prolem is given in Fig. The informtion signl x s recorded ith corrupting noise n tht s generted from nother noise source through n unknon nonliner process. The noise n is Gussin ith zero men nd unit vrince. The purpose of dptive noise cncelltion is to produce n nti-ve hose mgnitude is exctly the sme s tht of the unnted noise nd hose phse is exctly opposite. The scheme of dptive noise cncelltion system is shon s Fig.. The primry input source receives the desired signl s( k ) ith corrupting noise y( k ). The corrupting noise y( k) is generted y the noise source n( k ). The received signl of primry input is thus x( k) = s( k) + y( k ) (.) Signl source s(k) Noise n(k) source Σ y(k) Nonliner Chnnel x(k) = s(k) + y(k) FAMNN Filter ŷ(k) Σ ŝ(k) FAMNN noise cnceller Recovered signl Fig. The dptive noise cncelltion system A secondry (reference) input source receives noise n( k) uncorrelted ith the signl source s( k ) ut correlted ith the corrupting noise y( k ). This secondry input source provides the reference input to the dptive noise cnceller.the n( k) is used y n dptive process to generte n output yˆ( k ) (the output of the proposed FAMNNF) tht replic of y( k ). The output is then sutrcted from the primry input x( k ) to recover the desired signl s( k ). The sic ssumptions for the dptive noise cncelltion system include:. The s( k ), n( k) nd y( k ) re sttionry zeromen processes.. The s( k) is uncorrelted ith n( k) nd y( k ).. The n( k) is correlted ith y( k ).. The yˆ( k) re uncorrelted ith s( k ). From Fig., it follos sˆ ( k) = s( k) + n( k) yˆ ( k ) (.6) By squring eqution (), e hve sˆ ( k) = s ( k) + ( y( k) yˆ ( k)) + s( k)( y( k) yˆ ( k)) (.7) If e tke expecttion on oth sides, the folloing eqution is otined. E[ sˆ ( k)] = E[ s ( k)] + E[( y( k) yˆ ( k )) ] (.) The oective of dptive noise cncelltion is to minimize E[( y( k) yˆ ( k )) ],hich is equivlent to minimize E[ sˆ ( k )] y oserving eqution (.). If E[( y( k) yˆ ( k )) ] pproches to zero, then the remining error sˆ( k ) is exctly the sme s the desired signl s( k ). Thus, the signl E[( y( k) yˆ ( k)) ] ill e used s the error signl to turn the prmeters of the proposed FAMNN filter. Simultion Results Bse on the purposed techniques, commnd voices for the utomtic heel chir in Thi lnguge, hich re go hed, go ck, turn left, turn right nd stop, ere used to confirm the effectiveness. The mesurle output noisy speech ith db, db, -db, -db nd -db SNR level is tken s the contminted version of the interference, since the noise is not directly ville. The hite noise source is tken s the input. The simultion is shon in Fig. Fig.. Ech recovered signl re gree ith ech originl signl.

4 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, () Time, Sec Fig. Results of noise reduction for go hed in Thi lnguge, () Clen speech, () Noisy Speech ( db) nd Speech fter noise reduction () () ().... Fig. 6 Results of noise reduction for turn right in Thi lnguge, () Clen speech, () Noisy Speech (- db) nd Speech fter noise reduction. () () Time, (sec) Time, (sec) Fig. Results of noise reduction for go ck in Thi lnguge, () Clen speech, () Noisy Speech ( db) nd Speech fter noise reduction. () () Fig. 7 Results of noise reduction for turn left in Thi lnguge, () Clen speech, () Noisy Speech (- db) nd Speech fter noise reduction.

5 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, 7 6 () () Fig. Results of noise reduction for stop in Thi lnguge, () Clen speech, () Noisy Speech (- db) nd Speech fter noise reduction. Conclusion From the simultion results, the effort to enhnce speech ith FAMNN filtering scheme for noise cncelltion is cceptle. The utomtic heel chir orked ell hen using oth the originl nd recovered signls. The effectiveness of purposed noise reduction techniques s confirmed. Furthermore, the purposed noise reduction technique cn e pplied in rel ppliction. References: [] J. Thevril, H. K. Kn, Speech Enhncement using Adptive Neuro-Fuzzy Filtering, Proc. of Int. Sym. on Intelligent Signl Processing nd Communiction Systems, Decemer, pp []C. T. Lin, Single-chnnel Speech Enhncement in Vrile Noise-level Environment, IEEE Trns. on System, Mn nd Cyernetics Prt A, Vol., No., Jnury, pp. 7. [] T. W. Lee nd K. S. Yo, Speech Enhncement y Perceptul Filter ith Sequentil Noise Prmeter Estimtion Proc. of IEEE Int. Conf. on Acoustics, Speech nd Signl Processing, Vol., 7 My, pp []L. Singh nd S. Sridhrn, Speech enhncement using pre-processing, Proc. of IEEE Region Annul Conf. on Speech nd Imge Technologies for Computing nd Telecommunictions, Vol., Decemer 997, pp []J. D. Chen, J. Benesty, Y. T. Hung nd S. Doclo, Ne insights into the noise reduction Wiener filter, IEEE Trns. on Audio, Speech nd Lnguge Processing, Vol., No., July 6, pp.. [6]C. Plpous, C. Mrro, nd P. Sclrt, Improved Signl-to-Noise Rtio Estimtion for Speech Enhncement, IEEE Trns. on Audio, Speech nd Lnguge Processing, Vol., No.6, Novemer 6, pp. 9. [7]H. B. D. Sorensen, Speech recognition in noise using self-structuring noise reduction model nd hidden control models, Int. Joint Conf. on Neurl Netorks, Vol., 7- June 99, pp. 79. []N. Ssok, K. Sumi,; Y. Itoh nd K. Fuii, A ne noise reduction system sed on ALE nd noise reconstruction filter, IEEE Int. Sym. on Circuits nd Systems, Vol., My, pp [9]G. S. Kng nd D. A. Heide, Acoustic noise reduction for speech communiction: (secondorder grdient microphone) IEEE Int. Sym. on Circuits nd Systems, Vol., My- June 999, pp []K. Kroschel, Methods for noise reduction pplied to speech input systems, Proc. of VLSI nd Computer Peripherls. VLSI nd Microelectronic Applictions in Intelligent Peripherls nd their Interconnection Netorks, Vol., - My 99, pp. 7. [] G.A. Crpenter, S. Grosserg, N. Mrkuzon, J.H. Reynolds, nd D.B. Rosen, Fuzzy ARTMAP: A Neurl Netork Architecture for Incrementl Supervised Lerning of Anlog Multidimensionl Mps, IEEE Trns. on Neurl Netorks,Vol.,No.,99,pp []D. Chrlmpidis, T. Kspris nd M. Georgiopoulos, Clssifiction of noisy signls using fuzzy ARTMAP neurl netorks, IEEE Trns. on Neurl Netorks, Vol., No., Septemer, pp. 6. []S.Sumthi, S.Sivnndm nd N.Jgdeesri, Design of Soft Computing Models for Dt Mining Applictions, Indin Journl of Engineering nd Mterils Sciences, Vol. 7, No.,, pp. 7-.

6 Proceedings of the 7th WSEAS Interntionl Conference on Signl, Speech nd Imge Processing, Beiing, Chin, Septemer -7, 7 [] G.A. Crpenter, M.N. G, S. Gopl nd C.E. Woodcock, ART Neurl Netorks for Remote Sensing: Vegettion Clssifiction from Lndst TM nd Terrin Dt, IEEE Trns. on Geosciences nd Remote Sensing, Vol., No., 997. [] Murshed, N.A., Bortolozzi, F., nd Sourin R., A Cognitive Approch to Signture Verifiction. Interntionl Journl of Pttern Recognition nd Artificil Intelligence, Specil issue on Bnk Cheques Processing, Vol., No. 7, 997, pp.-, [6] Ty, Y. H., Khlid, M., Tn, K. K., nd Yusof, R., Hnd-Written Postcode Recognition y Fuzzy ARTMAP Neurl Netork, COSTAM Ntionl Science Congress, Genting Highlnds, Novemer, 997. [7] B. Mtthis, nd C. E. Sdik, Optoelectronic Fuzzy ARTMAP Processor, Journl of Opticl Computing, Vol., Mrch 99, pp. -. []C. Plpous, C. Mrro nd P. Sclrt, Improved Signl-to-Noise Rtio Estimtion for Speech Enhncement, IEEE Trns. on Audio, Speech nd Lnguge Processing, Vol., No. 6, Novemer 6, pp. 9. [9] S. N. Wu nd J. S. Wng. An dptive recurrent neuro-fuzzy filter for noisy speech enhncement Proc. of IEEE Int.Joint Conference on Neurl Netorks, Vol., -9 July, pp..