Deep Learning in Speech Synthesis by Google 2013

《Deep Learning in Speech Synthesis by Google 2013》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、DeepLearninginSpeechSynthesisHeigaZenGoogleAugust31st,2013OutlineBackgroundDeepLearningDeepLearninginSpeechSynthesisMotivationDeeplearning-basedapproachesDNN-basedstatisticalparametricspeechsynthesisExperimentsConclusionText-to-speechassequence-to-sequencemap

2、pingAutomaticspeechrecognition(ASR)Speech(continuoustimeseries)!Text(discretesymbolsequence)Machinetranslation(MT)Text(discretesymbolsequence)!Text(discretesymbolsequence)Text-to-speechsynthesis(TTS)Text(discretesymbolsequence)!Speech(continuoustimeseries)

3、HeigaZenDeepLearninginSpeechSynthesisAugust31st,20131of50Speechproductionprocesstext(concept)fundamentalreqvoiced/unoicedfreqtransercharfrequencyspeechtransfercharacteristicsmagnitudestart--endSoundsourcefundamentalvoiced:pulsefrequencyunvoiced:noisemodulatio

4、nofcarierwavebyspechinformationairflowHeigaZenDeepLearninginSpeechSynthesisAugust31st,20132of50Typical owofTTSsystemTEXTSentencesegmentaitonWordsegmentationTextnormalizationTextanalysisPart-of-speechtaggingPronunciationProsodypredictionSpeechsynthesisdiscrete

5、⇒discreteWaveformgenerationNLPFrontendSYNTHESIZEDdiscrete⇒continuousSPEECHSpeechBackendThistalkfocusesonbackendHeigaZenDeepLearninginSpeechSynthesisAugust31st,20133of50Statisticalparametricspeechsynthesis(SPSS)[2]FeatureModelParameterWaveformSynthesizedSpeech

6、extractiontraininggenerationsynthesisSpeechTextTextLargedata+automatictraining!AutomaticvoicebuildingParametricrepresentationofspeech!FlexibletochangeitsvoicecharacteristicsHiddenMarkovmodel(HMM)asitsacousticmodel!HMM-basedspeechsynthesissystem(HTS)[1]Heiga

7、ZenDeepLearninginSpeechSynthesisAugust31st,20134of50CharacteristicsofSPSSAdvantagesFlexibilitytochangevoicecharacteristicsSmallfootprintRobustnessDrawbackQualityMajorfactorsforqualitydegradation[2]VocoderAcousticmodel!DeeplearningOversmoothingHeigaZ

8、enDeepLearninginSpeechSynthesisAugust31st,20135of50Deeplearning[3]Machinelearningmethodologyusingmultiple-layeredmodelsMotivatedbybrains,whichorganizeideasandconceptshierarchicallyTypi