ACOUSTIC MODELING IN STATISTICAL PARAMETRIC SPEECH SYNTHESIS – FROM HMM TO LSTM-RNN英文学习材料

《ACOUSTIC MODELING IN STATISTICAL PARAMETRIC SPEECH SYNTHESIS – FROM HMM TO LSTM-RNN英文学习材料》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ACOUSTICMODELINGINSTATISTICALPARAMETRICSPEECHSYNTHESIS–FROMHMMTOLSTM-RNNHeigaZenGoogleABSTRACTStatisticalparametricspeechsynthesis(SPSS)combinesanacous-ticmodelandavocodertorenderspeechgivenatext.Typicallytextdecisiontree-clusteredcontext-dependenthiddenMarkovmodels(concept)(HMMs)

2、areemployedastheacousticmodel,whichrepresentarelationshipbetweenlinguisticandacousticfeatures.Recently,ar-tificialneuralnetwork-basedacousticmodels,suchasdeepneuralnetworks,mixturedensitynetworks,andlongshort-termmemoryrecurrentneuralnetworks(LSTM-RNNs),showedsignificantim-fundament

3、alreqvoiced/unoicedfreqtransercharprovementsovertheHMM-basedapproach.ThispaperreviewsfrequencyspeechtransfertheprogressofacousticmodelinginSPSSfromtheHMMtothecharacteristicsLSTM-RNN.magnitudestart--endIndexTerms—Statisticalparametricspeechsynthesis;artificialSoundsourcefundamentaln

4、euralnetworks;hiddenMarkovmodels;longshort-termmemory;voiced:pulsefrequencyunvoiced:noisemodulationofcarierwavebyspechinformation1.INTRODUCTIONairflowThegoaloftext-to-speech(TTS)synthesisistorenderanaturallyFig.1.Outlineofspeechproductionprocess.soundingspeechwaveformgivenatexttob

5、esynthesized.Figure1outlinesahumanspeechproductionprocess.Atext(orconcept)isfirsttranslatedintomovementsofarticulatorsandorgans.Usingback-endpartofTTSsystems,suchassmallfootprint[3,4]andflex-air-flowfromalung,vocalsourceexcitationsignalscontainingpe-ibilitytochangeitsvoicecharacteris

6、tics[5–8].However,thenatu-riodic(byvocalcordvibration)andaperiodic(byturbulentnoise)ralnessofthesynthesizedspeechfromSPSSisnotasgoodasthatcomponentsaregenerated.Byfilteringthesourcesignalsbytime-ofthebestsamplesfromconcatenativespeechsynthesizers.Zenetvaryingvocaltracttransferfunct

7、ionscontrolledbythearticulators,al.reportedthreemajorfactorsthatcandegradethenaturalness[1];theirfrequencycharacteristicsaremodulated.Finally,thefilteredqualityofvocoder,accuracyofacousticmodel,andeffectofover-sourcesignalsareemitted.TheaimofTTSistomimicthisprocesssmoothing.Thispap

8、eraddressestheaccuracyofacousticm