资源描述:
《语音识别_中英文献翻译》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、SpeechRecognitionVictorZue,RonCole,&WayneWardMITLaboratoryforComputerScience,Cambridge,Massachusetts,USAOregonGraduateInstituteofScience&Technology,Portland,Oregon,USACarnegieMellonUniversity,Pittsburgh,Pennsylvania,USA1DefiningtheProblemSpeechrecognitionistheprocessofconvertinganacousticsign
2、al,capturedbyamicrophoneoratelephone,toasetofwords.Therecognizedwordscanbethefinalresults,asforapplicationssuchascommands&control,dataentry,anddocumentpreparation.Theycanalsoserveastheinputtofurtherlinguisticprocessinginordertoachievespeechunderstanding,asubjectcoveredinsection.Speechrecognit
3、ionsystemscanbecharacterizedbymanyparameters,someofthemoreimportantofwhichareshowninFigure.Anisolated-wordspeechrecognitionsystemrequiresthatthespeakerpausebrieflybetweenwords,whereasacontinuousspeechrecognitionsystemdoesnot.Spontaneous,orextemporaneouslygenerated,speechcontainsdisfluencies,a
4、ndismuchmoredifficulttorecognizethanspeechreadfromscript.Somesystemsrequirespeakerenrollment---ausermustprovidesamplesofhisorherspeechbeforeusingthem,whereasothersystemsaresaidtobespeaker-independent,inthatnoenrollmentisnecessary.Someoftheotherparametersdependonthespecifictask.Recognitionisge
5、nerallymoredifficultwhenvocabulariesarelargeorhavemanysimilar-soundingwords.Whenspeechisproducedinasequenceofwords,languagemodelsorartificialgrammarsareusedtorestrictthecombinationofwords.Thesimplestlanguagemodelcanbespecifiedasafinite-statenetwork,wherethepermissiblewordsfollowingeachwordare
6、givenexplicitly.Moregenerallanguagemodelsapproximatingnaturallanguagearespecifiedintermsofacontext-sensitivegrammar.Onepopularmeasureofthedifficultyofthetask,combiningthevocabularysizeandthelanguagemodel,isperplexity,looselydefinedasthegeometricmeanofthenumberofwordsthatcanfollowawordafterthe
7、languagemodelhasbeenapplied(seesectionforadiscussionoflanguagemodelingingeneralandperplexityinparticular).Finally,therearesomeexternalparametersthatcanaffectspeechrecognitionsystemperformance,includingthecharacteristicsoftheenvironmentalnoise