搜索
359Vol.35,No.920099ACTAAUTOMATICASINICASeptember,2009DSP11,..,,,;,(A±neprojectionalgorithm,APA),,,.TMS320VC5509DSP.,(Recursiveleastsquares,RLS),(Leastmeansquares,LMS).,,,,TN912ResearchonAdaptiveSpeechEnhancementAlgorithmBasedonDi®erentialMicrophoneArrayandItsImplementationwithDSPSONGHui1LIUJia1AbstractAdaptive¯lteringisincommonuseinmostspeechenhancementalgorithms,whilecomplexityandconvergencespeedshouldbeconsidered¯rstwhenanadaptive¯lteringmethodisdesigned.Thispaperpresentsaspeechenhancementadaptive¯lteringmethodimplementedonchip.Thismethodisimplementedintwosteps.Inthe¯rststep,¯rst-orderdif-ferentialmicrophonearrayisutilizedtoobtainreal-timenoiseestimation.Inthesecondstep,thetraditionala±neprojectionalgorithm(APA)ismodi¯edsothatafasterrorvectorcalcu-lationmethodisobtained,andthesearchstepandprojectiondimensioncanbealtereddynamicallyaccordingtotheestima-tionerror.Thus,noisecanberemovedadaptively.ThismethodisrealizedonaTMS320VC5509DSPchip.ExperimentsshowthattheproposedmethodhasfastconvergencespeedlikeRLS(Recursiveleastsquares)andlowcomputationalcomplexitylikeLMS(Leastmeansquares).KeywordsSpeechenhancement,di®erentialmicrophonear-ray,a±neprojectionalgorithm(APA),adaptive¯lter,noisere-duction,.,2008-06-102009-01-19ReceivedJune10,2008;inrevisedformJanuary19,2009(863)(2006AA010101,2007AA04Z223)(60776800)SupportedbyNationalHighTechnologyResearchandDevelopmentProgramofChina(863Program)(2006AA010101,2007AA04Z223)andNationalNaturalScienceFoundationofChinaandMicrosoftResearchAsia(60776800)1.()1000841.TsinghuaNationalLaboratoryforInformationScienceandTech-nology,DepartmentofElectronicEngineering,TsinghuaUniversity,Beijing100084DOI:10.3724/SP.J.1004.2009.01240.,.,.,(Beamforming)[1],(Gener-alizedsidelobecanceller)[2],(Generalizedsingularvaluedecomposition)[3].,.,,,,,..,.(Leastmeansquares,LMS),,,,;(Recursiveleastsquares,RLS),LMS,,[4];(A±neprojectionalgorithm,APA)LMSRLS,,[5],,,.(Multi-stepmulti-dimensiona±neprojection,MMAP),.,.:1,\{;2MMAP,;3;4,MMAP;5.11.11.d.,{[6]:1Fig.1Diagramof¯rst-orderdi®erentialmicrophonearray9:DSP1241jH(!;µ)j=¯¯¯¯Y(!;µ)X(!)¯¯¯¯=¯¯¯1¡e¡j(!T+kd)¯¯¯=2sin!·T+(dcosµ)c¸2(1),c,d,k,k=!=c,µ,Y(!;µ)X(!)y(t)x(t).,,kd¿¼!T¿¼,(1)jH(!;µ)j=!·T+(dcosµ)c¸(2)2.2Fig.2Normalizedamplituderesponseof¯rst-orderdi®erentialmicrophonearray1.2,,.,,180±(Null),.,,,.,,.,\{,,,,,\.(2),{.,y(t)=x(t)¡x(t¡T)(3)H(!)=1¡e¡j!T.,,.,:HLFC(!)=1H(!)=11¡®e¡j!T(4).,®0!=0.2MMAP2.1APAeeen=yyyn¡XTn!!!¤n¡1(5)4!!!n=XnhXTnXn+±Ii¡1eee¤n(6)!!!n=!!!n¡1+¹4!!!n(7),!!!=[!0;!1;¢¢¢;!M¡1]TM,xxxn=[xxxn;xxxn¡1;¢¢¢;xxxn¡M+1]TnM,yyyn=[yyyn;yyyn¡1;¢¢¢;yyyn¡L+1]TL,L,Xn=[xxxn;xxxn¡1;¢¢¢;xxxn¡L+1]TM£L,eeenL,,¤,±I,,±,,¹.LM.L=1,APALMS;L¸MAPARLS.2.2LMS:¹;.,LMS,[4],.,APA,.,,;,,.L,APA.(5),:eeen=yyyn¡XTn!!!¤n¡1=yyyn¡xxxTn!!!¤n¡1~yyyn¡1¡~XTn¡1!!!¤n¡1#(8),~Xn¡1Xn¡1L¡1,XnL¡1,~yyyn¡1yyyn¡1L¡1.[7]:eeep;n¡1=yyyn¡1¡XTn¡1!!!¤n¡1=hI¡¹XTn¡1Xn¡1(XTn¡1Xn¡1+±I)¡1ieeen¡1(9),p\.(8)(9),L¡1L¡1.±,,:eeen=eeen(1¡¹)~eeen¡1#(10),(10)[7].,,.\,124235Lp:keeenkp=LXi=1jeeen;ijp#1p(11)Lp,.,;,.,keeenkp.,Lpkeeenkp=keeenk1.tttl=[tl1;tl2;¢¢¢]T,jtlij1;tlitlj8i;j;ij.,(0;1),,.,3.2.3.,¹,0¹2.,,.,:gm;n=keeemkp¡keeenkpm¡n(12),mn.,.gn¡c;n=g1;2,,c1.(12),.c-,.tttg=[tg1;tg2;¢¢¢]T,gn¡c;n0,,tgj1;tgitgj;8i;j;ij.,.Lp,,n,c.2.4,(Voiceactivitydetection,VAD),,\{.[8],,VAD..,VAD.3.3Fig.3Diagramofadaptivespeechenhancementsystem3DSP.(TI)TMS320VC5509DSP,PLLJTEGDMAGPIO(McBSPs).5509DSP,64K£16bitRAM64KBROM;ROM(Bootloader);,M5M29GB/T320VPFlash.2MW,128,(EMIF)DSP;;I/O,16MB;(ALU),200MHz,5ns,400MIPS.DSP4.4MMAPDSPFig.4DiagramofhardwarerealizationofMMAP4.d=3cm,8kHz,256,128..5,1Mos(Meanopinionscore).,SNR,RLS,RLS,,LMS4.MMAPRLS:Mos,.5MMAPFig.5ComparisonofspeechwaveformsandspectrogramsbeforeandafterMMAP9:DSP12431(dB)Mos(/5)Table1ImprovementofSNR(dBandMos(/5)SNRMMAPLMSAPARLS/MosSNR/MosSNR/MosSNR/MosSNR/Mos¡7.77/2.48.60/3.77.92/3.78.51/3.59.21/3.6¡2.40/2.913.98/3.913.50/3.913.87/3.814.37/3.93.91/3.219.58/3.918.88/3.819.38/4.020.65/4.09.97/3.724.20/4.123.78/4.124.13/4.125.03/4.115.42/4.027.46/4.227.30/4.127.41/4.227.93/4.1(a)(¹=0:2)(a)In°uenceofprojectiondimensiononconvergence(¹=0:2)(b)(LLL=[50202]T)(b)In°uenceofsearchsteponconvergence(LLL=[50202]T)(c)(¹=0:2)(c)Convergencecomparisonofdi®erentmethods(¹=0:2)6MMAPFig.6AnalysisofconvergenceperformanceofMMAP6,.,APALMS;APA,,,;,MMAP(RLS),RLS.APAMMAP,,,.,,MMAP,.(6),.7MMAPAPAXTnXn.,MMAPAPA,,MMAPAPA.,,,.7XTnXnFig.7ComparisonoflogconditionnumbersofXTnXn..APA,ML,7L2[7],4!!!nML,APA2ML+7L2;MMAP,(M+1)¹L+7¹L2,¹L;RLS4M2+4M+2.2,288MIPS,144MHzDSP.LMS,,.MMAPLLLtttl.,MMAPAPA,RLS.,MMAP.2Table2Comparisonofcomplexityandhardwaresimulationresultsofthreedi®erentadaptive¯lteringalgorithms(%)MMAP7:1£1030.1825.8APA1:3£1040.3230.4RLS2:6£1054.8849.01244355.\{,,,;,,(MMAP),.5509DSP.,MMAP,,.1»2,LMS,RLS.References1WeeS,ChenHW,YuZL.Self-calibration-basedro-bustnear-¯eldadaptivebeamformingformicrophonear-rays.IEEETransactionsonCircuitsandSystemsII:ExpressBriefs,2007,54(3):267¡2712WarsitzE,KruegerA,Haeb-UmbachR.Speechenhance-mentwithanewgeneralizedeigenvectorblockingmatrixforapplicationinageneralizedsidelobecanceller.In:Proc