汉语语音识别中区分性声调模型及最优集成方法.pdf

《汉语语音识别中区分性声调模型及最优集成方法.pdf》由会员上传分享，免费在线阅读，更多相关内容在应用文档-天天文库。

1、JournalofSoutheastUniversity(EnglishEdition)Vol.23,No.2,pp.174-178June2007ISSN1003—798DiscriminativetonemodeltrainingandoptimalintegrationforMandarinspeechrecognitionHuangHaoZhuJie(DepartmentofElectronicEngineering,ShanghaiJiaotongUniversity,Shanghai200240,China)Abstract:Twodiscriminativemethodsfors

2、olvingtoneproblemsinMandarinspeechrecognitionarepresented.First,discriminativetrainingontheHMM(hiddenMarkovmodel)basedtonemodelsisproposed.Thenanintegrationtechnigueoftonemodelsintoalargevocabularycontinuousspeechrecognitionsystemispresented.Discriminativemodelweighttrainingbasedonminimumphoneerrorc

3、riteriaisadoptedaimingatoptimalintegrationofthetonemodels.TheextendedBaumWelchalgorithmisappliedtofindthemodel-dependentweightstoscaletheacousticscoresandtonescores.Experimentalresultsshowthattonerecognitionratesandcontinuousspeechrecognitionaccuracycanbeimprovedbythediscriminativelytrainedtonemodel

4、.PerformanceofalargevocabularycontinuousMandarinspeechrecognitionsystemcanbefurtherenhancedbythediscriminativelytrainedweightcombinationsduetoabetterinterpolationofthegivenmodels.Keywords:discriminativetraining;minimumphoneerror;tonemodeling;MandarinspeechrecognitionTonerecognitionisanimportanttaskf

5、orManda-tonemodels,weproposeamethodtodiscriminativelyrinspeechrecognitionduetothetonalnatureoftheintegratetoneinformationintotheexistingsystems.Inlanguage.Therehasbeenmuchworkdonediscussingpreviousworkontonemodelingincorporationsuchastonemodelingtoimprovetonerecognitionaccuracy.inRefs.[23],aglobalac

6、ousticandtonemodelThemostpopularlyappliedistheframeandtheHMMweightiscommonlyappliedwhichmaynotobtainan[1][8]basedapproach.Inaddition,otherapproachessuchasoptimalresult.Liuetal.proposedanMCEbasedthestochasticpolynomialtonemodel(SPTM)pro-streamweightoptimiZationforaudio-visualLVSCR.posedinRef.[2]andth

7、edecisiontreebasedtoneInspiredbythiswork,weproposediscriminativetrain-[3]modelhavealsobeenproposed.ingonmodelweightusingtheEBalgorithmundertheInstate-of-the-artspeechrecognitionsystems,dis-MPEcriterio