资源描述：
《VOWEL DURATION MEASUREMENT USING DEEP NEURAL NETWORKS基于神经时间网络测定》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、2015IEEEINTERNATIONALWORKSHOPONMACHINELEARNINGFORSIGNALPROCESSING,SEPT.17–20,2015,BOSTON,USAVOWELDURATIONMEASUREMENTUSINGDEEPNEURALNETWORKSYossiAdi,JosephKeshetMatthewGoldrickDept.ofComputerScienceDept.ofLinguisticsBar-IlanUniversity,Ramat-Gan,IsraelNorthwesternUniversity,Evanston,IL,USAadiy
2、oss@cs.biu.ac.il,joseph.keshet@biu.ac.ilmatt-goldrick@northwestern.eduABSTRACTnetictranscriptionoftheinputsignal,andshouldbecarefullytuned[6].OurworkisfocusedoninputofspeechsegmentsVoweldurationsaremostoftenutilizedinstudiesaddressingwhereasinglevowelisprecededandfollowedbyconsonantspecificis
3、suesinphonetics.Thusfarthishasbeenhampered(CVC),andthephonetictranscriptionisnotneeded.byarelianceonsubjective,labor-intensivemanualannota-Herewetakeadifferentrouteandtrainaclassifieratthetion.Ourgoalistobuildanalgorithmforautomaticaccu-frame-leveltodetectwhethertheframeisavowelornot.Werateme
4、asurementofvowelduration,wheretheinputtotheusedstate-of-the-artdeepneuralnetwork(DNN)asaclassi-algorithmisaspeechsegmentcontainsonevowelprecededfier,comparingtwoDNNarchitectures:deepbeliefnetworkandfollowedbyconsonants(CVC).Ouralgorithmisbased(DBN)andconvolutionalneuralnetwork(CNN).Botharchi-
5、onadeepneuralnetworktrainedattheframelevelonman-tectureshaveproducedgoodresultsinpreviousspeechpro-uallyannotateddatafromaphoneticstudy.Specifically,wecessingstudies[7,8].Eacharchitecturewastrainedonman-trytwodeep-networkarchitectures:convolutionalneuralnet-uallyannotateddataandtheirperforman
6、cewascompared.Atwork(CNN),anddeepbeliefnetwork(DBN),andcompareinferencetime,theclassifierpredictstheprobabilityofeachtheiraccuracytoanHMM-basedforcedaligner.Resultssug-frameoftheinputasbeingavowel.ThepredictionsaregestthatCNNisbetterthanDBN,andbothCNNandHMM-smoothedouttohaveasinglechunkrepres
7、entingthevowel,basedforcedalignerarecomparableintheirresults,butnei-andthenvoweldurationiscomputed.therofthemyieldedthesamepredictionsasmodelsfittoWecomparetheaccuracyofDBN,CNNandHMM-manuallyannotateddata.basedforcealigneronmanuallyannotateddata.The
