语音识别系统毕业论文中英文资料对照外文翻译文献

语音识别中英文资料对照外文翻译文献SpeechRecognitionVictorZueRonColeampWayneWardMITLaboratoryforComputerScienceCambridgeMassachusettsUSAOregonGraduateInstituteofScienceampTechnologyPortlandOregonUSACarnegieMellonUniversityPittsburghPennsylvaniaUSA1DefiningtheProblemSpeechrecognitionistheprocessofconvertinganacousticsignalcapturedbyamicrophoneoratelephonetoasetofwords.Therecognizedwordscanbethefinalresultsasforapplicationssuchascommandsampcontroldataentryanddocumentpreparation.Theycanalsoserveastheinputtofurtherlinguisticprocessinginordertoachievespeechunderstandingasubjectcoveredinsection.SpeechrecognitionsystemscanbecharacterizedbymanyparameterssomeofthemoreimportantofwhichareshowninFigure.Anisolated-wordspeechrecognitionsystemrequires1thatthespeakerpausebrieflybetweenwordswhereasacontinuousspeechrecognitionsystemdoesnot.Spontaneousorextemporaneouslygeneratedspeechcontainsdisfluenciesandismuchmoredifficulttorecognizethanspeechreadfromscript.Somesystemsrequirespeakerenrollment---ausermustprovidesamplesofhisorherspeechbeforeusingthemwhereasothersystemsaresaidtobespeaker-independentinthatnoenrollmentisnecessary.Someoftheotherparametersdependonthespecifictask.Recognitionisgenerallymoredifficultwhenvocabulariesarelargeorhavemanysimilar-soundingwords.Whenspeechisproducedinasequenceofwordslanguagemodelsorartificialgrammarsareusedtorestrictthecombinationofwords.Thesimplestlanguagemodelcanbespecifiedasafinite-statenetworkwherethepermissiblewordsfollowingeachwordaregivenexplicitly.Moregenerallanguagemodelsapproximatingnaturallanguagearespecifiedintermsofacontext-sensitivegrammar.Onepopularmeasureofthedifficultyofthetaskcombiningthevocabularysizeandthelanguagemodelisperplexitylooselydefinedasthegeometricmeanofthenumberofwordsthatcanfollowawordafterthelanguagemodelhasbeenappliedseesectionforadiscussionoflanguagemodelingingeneralandperplexityinparticular.Finallytherearesomeexternalparametersthatcanaffectspeechrecognitionsystemperformanceincludingthecharacteristicsoftheenvironmentalnoiseandthetypeandtheplacementofthemicrophone.ParametersRangeSpeakingModeIsolatedwordstocontinuousspeechSpeakingStyleReadspeechtospontaneousspeechEnrollmentSpeaker-dependenttoSpeaker-independentVocabularySmalllt20wordstolargegt20000wordsLanguageModelFinite-statetocontext-sensitivePerplexitySmalllt10tolargegt100SNRHighgt30dBtolawlt10dBTransducerVoice-cancellingmicrophonetotelephoneTable:TypicalparametersusedtocharacterizethecapabilityofspeechrecognitionsystemsSpeechrecognitionisadifficultproblemlargelybecauseofthemanysourcesofvariabilityassociatedwiththesignal.Firsttheacousticrealizationsofphonemesthesmallestsoundunitsofwhichwordsarecomposedarehighlydependentonthecontextinwhichtheyappear.Thesephoneticvariabilitiesareexemplifiedbytheacousticdifferencesofthephoneme，Atwordboundariescontextualvariationscanbequitedramatic---makinggasshortagesoundlikegashshortageinAmericanEnglishanddevoandaresoundlikedevandareinItalian.Secondacousticvariabilitiescanresultfromchangesintheenvironmentaswellasinthepositionandcharacteristicsofthetransducer.Thirdwithin-speakervariabilitiescanresultfromchangesinthespeakersphysicalandemotionalstatespeakingrateorvoicequality.Finallydifferencesinsociolinguisticbackgrounddialectandvocaltractsizeandshapecancontributetoacross-speakervariabilities.Figureshowsthemajorcomponentsofatypicalspeechrecognitionsystem.Thedigitizedspeechsignalisfirsttransformedintoasetofusefulmeasurementsorfeaturesatafixedratetypicallyonceevery10--20msecseesectionsand11.3forsignalrepresentationanddigitalsignalprocessingrespectively.Thesemeasurementsarethenusedtosearchforthemostlikelywordcandidatemakinguseofconstraintsimposedbytheacousticlexicalandlanguagemodels.Throughoutthisprocesstrainingdataareusedtodeterminethevaluesofthemodelparameters.Figure:Componentsofatypicalspeechrecognitionsystem.Speechrecognitionsystemsattempttomodelthesourcesofvariabilitydescribedaboveinseveralways.Atthelevelofsignalrepresentationresearchershavedevelopedrepresentationsthatemphasizeperceptuallyimportantspeaker-independentfeaturesofthesignalandde-emphasizespeaker-dependentcharacteristics.Attheacousticphoneticlevelspeakervariabilityistypicallymodeledusingstatisticaltechniquesappliedtolargeamountsofdata.Speakeradaptationalgorithmshavealsobeendevelopedthatadaptspeaker-independentacousticmodelstothoseofthecurrentspeakerduringsystemuseseesection.Effectsoflinguisticcontextattheacousticphoneticlevelaretypicallyhandledbytrainingseparatemodelsforphonemesindifferentcontextsthisiscalledcontextdependentacousticmodeling.Wordlevelvariabilitycanbehandledbyallowingalternatepronunciationsofwordsinrepresentationsknownaspronunciationnetworks.Commonalternatepronunciationsofwordsaswellaseffectsofdialectandaccentarehandledbyallowingsearchalgorithmstofindalternatepathsofphonemesthroughthesenetworks.Statisticallanguagemodelsbasedonestimatesofthefrequencyofoccurrenceofwordsequencesareoftenusedtoguidethesearchthroughthemostprobablesequenceofwords.ThedominantrecognitionparadigminthepastfifteenyearsisknownashiddenMarkovmodelsHMM.AnHMMisadoublystochasticmodelinwhichthegenerationoftheunderlyingphonemestringandtheframe-by-framesurfaceacousticrealizationsarebothrepresentedprobabilisticallyasMarkovprocessesasdiscussedinsectionsand11.2.NeuralnetworkshavealsobeenusedtoestimatetheframebasedscoresthesescoresarethenintegratedintoHMM-basedsystemarchitecturesinwhathascometobeknownashybridsystemsasdescribedinsection11.5.Aninterestingfeatureofframe-basedHMMsystemsisthatspeechsegmentsareidentifiedduringthesearchprocessratherthanexplicitly.Analternateapproachistofirstidentifyspeechsegmentsthenclassifythesegmentsandusethesegmentscorestorecognizewords.Thisapproachhasproducedcompetitiverecognitionperformanceinseveraltasks.2StateoftheArtCommentsaboutthestate-of-the-artneedtobemadeinthecontextofspecificapplicationswhichreflecttheconstraintsonthetask.Moreoverdifferenttechnologiesaresometimesappropriatefordifferenttasks.Forexamplewhenthevocabularyissmalltheentirewordcanbemodeledasasingleunit.Suchanapproachisnotpracticalforlargevocabularieswherewordmodelsmustbebuiltupfromsubwordunits.PerformanceofspeechrecognitionsystemsistypicallydescribedintermsofworderrorrateEdefinedas:whereNisthetotalnumberofwordsinthetestsetandSIandDarethetotalnumberofsubstitutionsinsertionsanddeletionsrespectively.Thepastdecadehaswitnessedsignificantprogressinspeechrecognitiontechnology.Worderrorratescontinuetodropbyafactorof2everytwoyears.Substantialprogresshasbeenmadeinthebasictechnologyleadingtotheloweringofbarrierstospeakerindependencecontinuousspeechandlargevocabularies.Thereareseveralfactorsthathavecontributedtothisrapidprogress.FirstthereisthecomingofageoftheHMM.HMMispowerfulinthatwiththeavailabilityoftrainingdatatheparametersofthemodelcanbetrainedautomaticallytogiveoptimalperformance.Secondmuchefforthasgoneintothedevelopmentoflargespeechcorporaforsystemdevelopmenttrainingandtesting.Someofthesecorporaaredesignedforacousticphoneticresearchwhileothersarehighlytaskspecific.Nowadaysitisnotuncommontohavetensofthousandsofsentencesavailableforsystemtrainingandtesting.Thesecorporapermitresearcherstoquantifytheacousticcuesimportantforphoneticcontrastsandtodetermineparametersoftherecognizersinastatisticallymeaningfulway.Whilemanyofthesecorporae.g.TIMITRMATISandWSJseesection12.3wereoriginallycollectedunderthesponsorshipoftheU.S.DefenseAdvancedResearchProjectsAgencyARPAtospurhumanlanguagetechnologydevelopmentamongitscontractorstheyhaveneverthelessgainedworld-wideacceptancee.g.inCanadaFranceGermanyJapanandtheU.K.asstandardsonwhichtoevaluatespeechrecognition.Thirdprogresshasbeenbroughtaboutbytheestablishmentofstandardsforperformanceevaluation.Onlyadecadeagoresearcherstrainedandtestedtheirsystemsusinglocallycollecteddataandhadnotbeenverycarefulindelineatingtrainingandtestingsets.Asaresultitwasverydifficulttocompareperformanceacrosssystemsandasystemsperformancetypicallydegradedwhenitwaspresentedwithpreviouslyunseendata.Therecentavailabilityofalargebodyofdatainthepublicdomaincoupledwiththespecificationofevaluationstandardshasresultedinuniformdocumentationoftestresultsthuscontributingtogreaterreliabilityinmonitoringprogresscorpusdevelopmentactivitiesandevaluationmethodologiesaresummarizedinchapters12and13respectively.Finallyadvancesincomputertechnologyhavealsoindirectlyinfluencedourprogress.Theavailabilityoffastcomputerswithinexpensivemassstoragecapabilitieshasenabledresearcherstorunmanylargescaleexperimentsinashortamountoftime.Thismeansthattheelapsedtimebetweenanideaanditsimplementationandevaluationisgreatlyreduced.Infactspeechrecognitionsystemswithreasonableperformancecannowruninrealtimeusinghigh-endworkstationswithoutadditionalhardware---afeatunimaginableonlyafewyearsago.OneofthemostpopularandpotentiallymostusefultaskswithlowperplexityPP11istherecognitionofdigits.ForAmericanEnglishspeaker-independentrecognitionofdigitstringsspokencontinuouslyandrestrictedtotelephonebandwidthcanachieveanerrorrateof0.3whenthestringlengthisknown.Oneofthebestknownmoderate-perplexitytasksisthe1000-wordso-calledResourceManagementRMtaskinwhichinquiriescanbemadeconcerningvariousnavalvesselsinthePacificocean.Thebestspeaker-independentperformanceontheRMtaskislessthan4usingaword-pairlanguagemodelthatconstrainsthepossiblewordsfollowingagivenwordPP60.Morerecentlyresearchershavebeguntoaddresstheissueofrecognizingspontaneouslygeneratedspeech.ForexampleintheAirTravelInformationServiceATISdomainworderrorratesoflessthan3hasbeenreportedforavocabularyofnearly2000wordsandabigramlanguagemodelwithaperplexityofaround15.Highperplexitytaskswithavocabularyofthousandsofwordsareintendedprimarilyforthedictationapplication.Afterworkingonisolated-wordspeaker-dependentsystemsformanyyearsthecommunityhassince1992movedtowardsvery-large-vocabulary20000wordsandmorehigh-perplexityPP≈200speaker-independentcontinuousspeechrecognition.Thebestsystemin1994achievedanerrorrateof7.2onreadsentencesdrawnfromNorthAmericabusinessnews.Withthesteadyimprovementsinspeechrecognitionperformancesystemsarenowbeingdeployedwithintelephoneandcellularnetworksinmanycountries.Withinthenextfewyearsspeechrecognitionwillbepervasiveintelephonenetworksaroundtheworld.Therearetremendousforcesdrivingthedevelopmentofthetechnologyinmanycountriestouchtonepenetrationislowandvoiceistheonlyoptionforcontrollingautomatedservices.Invoicedialingforexampleuserscandial10--20telephonenumbersbyvoicee.g.callhomeafterhavingenrolledtheirvoicesbysayingthewordsassociatedwithtelephonenumbers.ATampTontheotherhandhasinstalledacallroutingsystemusingspeaker-independentword-spottingtechnologythatcandetectafewkeyphrasese.g.persontopersoncallingcardinsentencessuchas:Iwanttochargeittomycallingcard.Atpresentseveralverylargevocabularydictationsystemsareavailablefordocumentgeneration.Thesesystemsgenerallyrequirespeakerstopausebetweenwords.Theirperformancecanbefurtherenhancedifonecanapplyconstraintsofthespecificdomainsuchasdictatingmedicalreports.Eventhoughmuchprogressisbeingmademachinesarealongwayfromrecognizingconversationalspeech.WordrecognitionratesontelephoneconversationsintheSwitchboardcorpusarearound50.Itwillbemanyyearsbeforeunlimitedvocabularyspeaker-independentcontinuousdictationcapabilityisrealized.3FutureDirectionsIn1992theU.S.NationalScienceFoundationsponsoredaworkshoptoidentifythekeyresearchchallengesintheareaofhumanlanguagetechnologyandtheinfrastructureneededtosupportthework.Thekeyresearchchallengesaresummarizedin.Researchinthefollowingareasforspeechrecognitionwereidentified:Robustness:Inarobustsystemperformancedegradesgracefullyratherthancatastrophicallyasconditionsbecomemoredifferentfromthoseunderwhichitwastrained.Differencesinchannelcharacteristicsandacousticenvironmentshouldreceiveparticularattention.Portability:Portabilityreferstothegoalofrapidlydesigningdevelopinganddeployingsystemsfornewapplications.Atpresentsystemstendtosuffersignificantdegradationwhenmovedtoanewtask.Inordertoreturntopeakperformance