文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Continuous Speech Recognition at
侵权投诉

Continuous Speech Recognition at

Abstract: This paper summarizes recent activities at LIMSI in multilingual speech recognition and its applications. While the main goal of speech recognition is to provide a transcription of the speech signal as a sequence of words, the same basic technolo

RecentActivitiesinSpokenLanguageProcessingatLIMSI

Jean-LucGauvainandLoriLamel

SpokenLanguageProcessingGroup,LIMSI-CNRS,91403Orsay,FRANCE

gauvain,lamel@limsi.frdoc.guandang.net

Abstract:ThispapersummarizesrecentactivitiesatLIMSIinmultilingualspeechrecognitionanditsapplications.Whilethemaingoalofspeechrecognitionistoprovideatranscriptionofthespeechsignalasasequenceofwords,thesamebasictechnologyservesasthe rststepinotherapplicationareas,suchasinautomaticsystemsforinformationaccessandforautomaticindexationofaudiovisualdata.

SPEECHRECOGNITION

Speechrecognitionisprincipallyconcernedwiththeproblemoftranscribingthespeechsignalasasequenceofwords.TheLIMSIsystem,incommonwithmostoftoday’sstate-of-the-artsystems(4),makesuseofstatisticalmodelsofspeechgeneration.Fromthispointofview,messagegenerationisrepresentedbyalanguagemodelwhichprovidesanestimateoftheprobabilityofanygivenwordstring,andtheencodingofthemessageintheacousticsignalisrepresentedbyaprobabilitydensityfunction(HMM).Thespeechdecodingproblemthenconsistsofmaximizingtheaposterioriprobabilityofthewordstringgiventheobservedacousticsignal.

TheLIMSIwordrecognizer(makesuseofcontinuousdensityHMMsforacousticmodelingofcontextualphoneunitsandn-gramstatisticsestimatedontrainingtexts(newspapertextsand/orspeechtranscriptions)forlanguagemodeling.AcousticmodelingusesPLP-likecepstralfeaturesderivedfromaMelfrequencyspectrum,withsentence-basedcepstralnormalization.Foreachtasktherecognitionvocabularyisselectedtomaximizelexicalcoverage,thusminimizingerrorsduetounknownwords.Lexicalpronunciationsarephonemically-based,withlanguage-speci csymbols(7).Speci cunitsareusedtomodelnon-speecheffectssuchassilence, llerwords,andbreathnoises.Fortranscriptiontasks,wordrecognitioniscarriedoutinmultipledecodingpasses,wheretheinformationbetweenlevelsistransmittedviawordgraphs(5).Inthe rstpassawordgraphisgeneratedusingasmallbigramlanguagemodel,whichisthenusedtoconstrainthesearchspaceinfurtherdecodingpasseswithmoreaccurateacousticandlanguagemodels.Unsupervisedacousticmodeladaptationmayoptionallybeperformedtoimproveperformance.TheapplicabilityofspeechrecognitiontechniquesfordifferentlanguagesisofparticularimportanceinEurope.AtLIMSI,speaker-independent,largevocabulary,continuousspeechrecognitionsystemsfordifferentEuropeanlan-guages(French,GermanandBritishEnglish)andforAmericanEnglishhavebeendeveloped(9).ForAmericanEnglishworderrorratesofabout8%canbeachievedforunrestrictednewspapertextsfromunknownspeaker.WorderrorsonasimilartaskinFrenchareabout11%(1).Thewidelyusedn-gramlanguagemodelsarereasonablysuc-cessfulfordictationinEnglish,butarelessef cientformorehighlyin ectedlanguagessuchasFrenchandGerman.Applicationsofthistechnologyaremainlyinprofessionaldomains,wherethesystemshandleaspokenencodingofwrittenlanguage.

TRANSCRIPTIONOFRADIOANDTELEVISIONBROADCASTS

Recentadvancesintheunderlyingspeechrecognitiontechnologyhaveledtoactivityusingreal-world(or“found”)speechdataincontrasttospeechproducedwiththeaimofbeingrecognizedbyamachine.Oneparticularlyexcitingareaofresearchisthetranscriptionoftelevisionandradiobroadcastsforwhichautomaticprocessingcaneasetheworkloadrequiredforindexationandretrievalpurposes(2).Thistaskischallengingfromthetechnicalview,astheshowscontaindataofvariousacousticandlinguisticnature,withabruptorgradualtransitionsbetweensegments.Thesignaltypicallyconsistsofpreparedspeechandspontaneousspeech,whichmaybeofstudioqualityorhavebeentransmittedoveratelephoneorothernoisychannel(ie.,corruptedbyadditivenoiseandnonlineardistorsions),aswellasspeechovermusicandpuremusicsegments.Acousticmodelstrainedoncleanspeechareclearlyinadequatetoprocesssuchinhomogeneousdata.Thespeechisproducedbyawidevarietyofspeakers:newsanchorsandtalkshowhosts,reportersinremotelocations,interviewswithpoliticiansandcommonpeople,unknownspeakers,newdialects,non-nativespeakers,etc.

Inadditiontoproblemsassociatedwithspeechtranscription,thecontinuousstreamofdataneedstobepartitionedintohomogeneoussegmentsinordertoachieveacceptableperformance(6).Fromrecentresearchitappearsthatcurrentwordrecognitionperformanceisnotcriticallydependentuponthepartitioningaccuracyandthatanyreason-ableapproachthatseparatesspeakerturnsandmajoracousticboundariesissuf cient.Inordertoaddressvariability

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email