语音自适应和差分编码方法

1SpeechSignalProcessing—Lecture7SpeechCodingMethodsBasedonSpeechWaveformRepresentationsandSpeechModels—AdaptiveandDifferentialCoding2QuantizationDilemmaWanttochoosequantizationstepsizelargeenoughtoaccommodatemaximumpeak-to-peakrangeofx(n);atthesametimeneedtomakethequantizationstepsizesmallsoastominimizethequantizationerror–thenon-stationarynatureofspeech(variabilityacrosssounds,speakers,backgrounds)compoundsthisproblemgreatly3SolutionstoQuantizationDilemma•Solution1-letΔvarytomatchthevarianceoftheinputsignal=Δ(n)•Solution2-useavariablegain,G(n),followedbyafixedquantizerstepsize,Δ=keepsignalvarianceofy(n)=G(n)x(n)constantCase1:Δ(n)proportionaltoσx2=quantizationlevelsandrangeswouldbelinearlyscaledtomatchσx2=needtoreliablyestimateσx2Case2:G(n)proportionalto1/σx2togiveσy2≈constant•needreliableestimateofσx2forbothtypesofadaptivequantizationAdaptiveQuantization:4TypesofAdaptiveQuantizationClassification1:•instantaneous-amplitudechangesreflectsample-to-samplevariationsinx(n)=rapidadaptation•syllabic-amplitudechangesreflectsyllable-to-syllablevariationsinx(n)=slowadaptationClassification2:•feed-forward-adaptivequantizersthatestimateσx2fromx(n)itself•feedback-adaptivequantizersthatadaptthestepsize,Δ,onthebasisofthequantizedsignal,,(orequivalentlythecodewords,c(n)))(ˆnx5FeedForwardAdaptationVariablestepsize•assumeuniformquantizerwithstepsizeΔ(n)•x(n)isquantizedusingΔ(n)=c(n)andΔ(n)needtobetransmittedtothedecoder•ifc’(n)=c(n)andΔ’(n)=Δ(n)=noerrorsinchannel,and)(ˆ)(ˆnxnxDon’thavex(n)atthedecodertoestimateΔ(n)=needtotransmitΔ(n);thisisamajordrawbackoffeedforwardadaptation6Feed-ForwardQuantizertimevaryinggain,G(n)=c(n)andG(n)needtobetransmittedtothedecoderCan’testimateG(n)atthedecoder=ithastobetransmitted7FeedForwardQuantizers•feedforwardsystemsmakeestimatesofσx2,thenmakeΔorthequantizationlevelsproportionaltoσx,orthegainisinverselyproportionaltoσx222221221assumeshort-timeenergy()()()where()isalowpassfilter()(thiscanbeshown)consider()10()()0xmxnnmnxmhnmhnEnhnnotherwisenxm1222001()(1)(1)(recursion)thisgives()and()/()nmnnxn(n)nGnGn8FeedForwardQuantizer•theparameterαcontrolstheeffectiveintervalofx(n)thatcontributestotheestimateof)(n2α=0.99=bringsuplevelinlowamplituderegions=syllabicrateα=0.9=systemreactstoamplitudevariationsmorerapidly=instantaneousrate9FeedForwardQuantizers•Δ(n)andG(n)varyslowlycomparedtox(n)–theymustbesampledandtransmittedaspartofthewaveformcoderparameters–rateofsamplingdependsonthebandwidthofthelowpassfilter,h(n)—forα=0.99,therateisabout13Hz;forα=0.9,therateisabout135Hzminmaxminmaxmaxminitisreasonabletoplacelimitsonthevariationof()or(),oftheform()()forobtainingconstantSNRovera40dBrangeinsignallevelsnGnGGnGnGGmaxmin100(40dBrange)10FeedForwardAdaptationGain1221()()()or()evaluatedeverysamplesused128,1024samplesforestimatesadaptivequantizerachievesupto5.6dBbetterSNRthannon-adaptivequantizerscanacnMmnnxmMnGnMMminmaxhievethisSNRwithlowidlechannelnoiseandwidespeechdynamicrangebysuitablechoiceofandlessgainforM=1024thanM=128by3dB=M=1024istoolonganinterval11FeedbackAdaptation•σ2(n)estimatedfromquantizeroutput(orthecodewords)•advantageoffeedbackadaptationisthatneitherΔ(n)norG(n)needstobetransmittedtothedecodersincetheycanbederivedfromthecodewords•disadvantageoffeedbackadaptationisincreasedsensitivitytoerrorsincodewords,sincesucherrorsaffectΔ(n)andG(n)12FeedbackAdaptation2221122ˆ()()()ˆ()basedonlyonpastvaluesof()twotypicalwindows/filtersare1.()102.()1/101ˆ()()canusevmnnmnMnxmhnmnxnhnnotherwisehnMnMotherwisenxmMeryshortwindowlengths(e.g.,2)toachieve12dBSNRfora3bitquantizerMB13PerformanceComparisonofAdaptiveQuantizationandPCM•improvementsinSNR•4-7dBimprovementoverμ-law14DifferentialQuantization•wehavecarriedinstantaneousquantizationofx(n)asfaraspossible•timetoconsidercorrelationsbetweenspeechsamplesseparatedintime=differentialquantization•highcorrelationvalues=signaldoesnotchangerapidlyintime=differencebetweenadjacentsamplesshouldhavelowervariancethanthesignalitselfdifferentialquantizationcanincreaseSNRatagivenbitrate,orlowerbitrateforagivenSNR15DifferentialQuantization()()()where()unquantizedinputsampleestimateorpredictionof()istheoutputofapredictorsystem,,ˆwhoseinputis(),aquantizedversionodnxnxnxnx(n)xnx(n)Pxnf()()predictionerrorsignalˆquantizeddifference(predictionerror)signalxndnd(n)16DifferentialQuantization•differencesignal,d(n),isquantized-notx(n)•quantizercanbefixed,oradaptive,uniformornon-uniform•quantizerparametersareadjustedtomatchthevarianceofd(n)ˆ()()()()quantizationerrorˆˆ()()()predictedplusquantizedˆ()()()quantizedinputhassamequantizationerrorasthedifferencesignalifddndnenenxnxndnxdxnxnen22,errorissmallerindependentofpredictor,,quantizeddiffersfromunquantizedby,thequantizationerrorofthedifferencesignal!goodpredictiongiveslowerquantizationxPx(n)x(n)e(n)errorthanquantizinginputdirectly17DifferentialQuantization•quantizeddifferencesignalisencodedintoc(