通信103屠建国翻译1Word下载.docx

通信103屠建国翻译1Word下载.docx

《通信103屠建国翻译1Word下载.docx》由会员分享，可在线阅读，更多相关《通信103屠建国翻译1Word下载.docx（12页珍藏版）》请在冰豆网上搜索。

MITLaboratoryforComputerScience,Cambridge,Massachusetts,USA

OregonGraduateInstituteofScience&

Technology,Portland,Oregon,USA

CarnegieMellonUniversity,Pittsburgh,Pennsylvania,USA

1DefiningtheProblem

Speechrecognitionistheprocessofconvertinganacousticsignal,capturedbyamicrophoneoratelephone,toasetofwords.Therecognizedwordscanbethefinalresults,asforapplicationssuchascommands&

control,dataentry,anddocumentpreparation.Theycanalsoserveastheinputtofurtherlinguisticprocessinginordertoachievespeechunderstanding,asubjectcoveredinsection?

.

Speechrecognitionsystemscanbecharacterizedbymanyparameters,someofthemoreimportantofwhichareshowninFigure?

.Anisolated-wordspeechrecognitionsystemrequiresthatthespeakerpausebrieflybetweenwords,whereasacontinuousspeechrecognitionsystemdoesnot.Spontaneous,orextemporaneouslygenerated,speechcontainsdisfluencies,andismuchmoredifficulttorecognizethanspeechreadfromscript.Somesystemsrequirespeakerenrollment---ausermustprovidesamplesofhisorherspeechbeforeusingthem,whereasothersystemsaresaidtobespeaker-independent,inthatnoenrollmentisnecessary.Someoftheotherparametersdependonthespecifictask.Recognitionisgenerallymoredifficultwhenvocabulariesarelargeorhavemanysimilar-soundingwords.Whenspeechisproducedinasequenceofwords,languagemodelsorartificialgrammarsareusedtorestrictthecombinationofwords.

Thesimplestlanguagemodelcanbespecifiedasafinite-statenetwork,wherethepermissiblewordsfollowingeachwordaregivenexplicitly.Moregenerallanguagemodelsapproximatingnaturallanguagearespecifiedintermsofacontext-sensitivegrammar.

Onepopularmeasureofthedifficultyofthetask,combiningthevocabularysizeandthelanguagemodel,isperplexity,looselydefinedasthegeometricmeanofthenumberofwordsthatcanfollowawordafterthelanguagemodelhasbeenapplied（seesection?

foradiscussionoflanguagemodelingingeneralandperplexityinparticular）.Finally,therearesomeexternalparametersthatcanaffectspeechrecognitionsystemperformance,includingthecharacteristicsoftheenvironmentalnoiseandthetypeandtheplacementofthemicrophone.

Parameters

Range

SpeakingMode

Isolatedwordstocontinuousspeech

SpeakingStyle

Readspeechtospontaneousspeech

Enrollment

Speaker-dependenttoSpeaker-independent

Vocabulary

Small（<

20words）tolarge（>

20,000words）

LanguageModel

Finite-statetocontext-sensitive

Perplexity

10）tolarge（>

100）

SNR

High（>

30dB）tolaw（<

10dB）

Transducer

Voice-cancellingmicrophonetotelephone

Table:

Typicalparametersusedtocharacterizethecapabilityofspeechrecognitionsystems

Speechrecognitionisadifficultproblem,largelybecauseofthemanysourcesofvariabilityassociatedwiththesignal.First,theacousticrealizationsofphonemes,thesmallestsoundunitsofwhichwordsarecomposed,arehighlydependentonthecontextinwhichtheyappear.Thesephoneticvariabilitiesareexemplifiedbytheacousticdifferencesofthephoneme，Atwordboundaries,contextualvariationscanbequitedramatic---makinggasshortagesoundlikegashshortageinAmericanEnglish,anddevoandaresoundlikedevandareinItalian.

Second,acousticvariabilitiescanresultfromchangesintheenvironmentaswellasinthepositionandcharacteristicsofthetransducer.Third,within-speakervariabilitiescanresultfromchangesinthespeaker'

sphysicalandemotionalstate,speakingrate,orvoicequality.Finally,differencesinsociolinguisticbackground,dialect,andvocaltractsizeandshapecancontributetoacross-speakervariabilities.

Figureshowsthemajorcomponentsofatypicalspeechrecognitionsystem.Thedigitizedspeechsignalisfirsttransformedintoasetofusefulmeasurementsorfeaturesatafixedrate,typicallyonceevery10--20msec（seesectionsand11.3forsignalrepresentationanddigitalsignalprocessing,respectively）.Thesemeasurementsarethenusedtosearchforthemostlikelywordcandidate,makinguseofconstraintsimposedbytheacoustic,lexical,andlanguagemodels.Throughoutthisprocess,trainingdataareusedtodeterminethevaluesofthemodelparameters.

Figure:

Componentsofatypicalspeechrecognitionsystem.

Speechrecognitionsystemsattempttomodelthesourcesofvariabilitydescribedaboveinseveralways.Atthelevelofsignalrepresentation,researchershavedevelopedrepresentationsthatemphasizeperceptuallyimportantspeaker-independentfeaturesofthesignal,andde-emphasizespeaker-dependentcharacteristics.Attheacousticphoneticlevel,speakervariabilityistypicallymodeledusingstatisticaltechniquesappliedtolargeamountsofdata.Speakeradaptationalgorithmshavealsobeendevelopedthatadaptspeaker-independentacousticmodelstothoseofthecurrentspeakerduringsystemuse,（seesection）.Effectsoflinguisticcontextattheacousticphoneticlevelaretypicallyhandledbytrainingseparatemodelsforphonemesindifferentcontexts;

thisiscalledcontextdependentacousticmodeling.

Wordlevelvariabilitycanbehandledbyallowingalternatepronunciationsofwordsinrepresentationsknownaspronunciationnetworks.Commonalternatepronunciationsofwords,aswellaseffectsofdialectandaccentarehandledbyallowingsearchalgorithmstofindalternatepathsofphonemesthroughthesenetworks.Statisticallanguagemodels,basedonestimatesofthefrequencyofoccurrenceofwordsequences,areoftenusedtoguidethesearchthroughthemostprobablesequenceofwords.

ThedominantrecognitionparadigminthepastfifteenyearsisknownashiddenMarkovmodels（HMM）.AnHMMisadoublystochasticmodel,inwhichthegenerationoftheunderlyingphonemestringandtheframe-by-frame,surfaceacousticrealizationsarebothrepresentedprobabilisticallyasMarkovprocesses,asdiscussedinsections,and11.2.Neuralnetworkshavealsobeenusedtoestimatetheframebasedscores;

thesescoresarethenintegratedintoHMM-basedsystemarchitectures,inwhathascometobeknownashybridsystems,asdescribedinsection11.5.

Aninterestingfeatureofframe-basedHMMsystemsisthatspeechsegmentsareidentifiedduringthesearchprocess,ratherthanexplicitly.Analternateapproachistofirstidentifyspeechsegments,thenclassifythesegmentsandusethesegmentscorestorecognizewords.Thisapproachhasproducedcompetitiverecognitionperformanceinseveraltasks.

2StateoftheArt

Commentsaboutthestate-of-the-artneedtobemadeinthecontextofspecificapplicationswhichreflecttheconstraintsonthetask.Moreover,differenttechnologiesaresometimesappropriatefordifferenttasks.Forexample,whenthevocabularyissmall,theentirewordcanbemodeledasasingleunit.Suchanapproachisnotpracticalforlargevocabularies,wherewordmodelsmustbebuiltupfromsubwordunits.

PerformanceofspeechrecognitionsystemsistypicallydescribedintermsofworderrorrateE,definedas:

whereNisthetotalnumberofwordsinthetestset,andS,I,andDarethetotalnumberofsubstitutions,insertions,anddeletions,respectively.

Thepastdecadehaswitnessedsignificantprogressinspeechrecognitiontechnology.Worderrorratescontinuetodropbyafactorof2everytwoyears.Substantialprogresshasbeenmadeinthebasictechnology,leadingtotheloweringofbarrierstospeakerindependence,continuousspeech,andlargevocabularies.Thereareseveralfactorsthathavecontributedtothisrapidprogress.First,thereisthecomingofageoftheHMM.HMMispowerfulinthat,withtheavailabilityoftrainingdata,theparametersofthemodelcanbetrainedautomaticallytogiveoptimalperformance.

Second,muchefforthasgoneintothedevelopmentoflargespeechcorporaforsystemdevelopment,training,andtesting.Someofthesecorporaaredesignedforacousticphoneticresearch,whileothersarehighlytaskspecific.Nowadays,itisnotuncommontohavetensofthousandsofsentencesavailableforsystemtrainingandtesting.Thesecorporapermitresearcherstoquantifytheacousticcuesimportantforphoneticcontrastsandtodetermineparametersoftherecognizersinastatisticallymeaningfulway.Whilemanyofthesecorpora（e.g.,TIMIT,RM,ATIS,andWSJ;

seesection12.3）wereoriginallycollectedunderthesponsorshipoftheU.S.DefenseAdvancedResearchProjectsAgency（ARPA）tospurhumanlanguagetechnologydevelopmentamongitscontractors,theyhaveneverthelessgainedworld-wideacceptance（e.g.,inCanada,France,Germany,Japan,andtheU.K.）asstandardsonwhichtoevaluatespeechrecognition.

Third,progresshasbeenbroughtaboutbytheestablishmentofstandardsforperformanceevaluation.Onlyadecadeago,researcherstrainedandtestedtheirsystemsusinglocallycollecteddata,andhadnotbeenverycarefulindelineatingtrainingandtestingsets.Asaresult,itwasverydifficulttocompareperformanceacrosssystems,andasystem'

sperformancetypicallydegradedwhenitwaspresentedwithpreviouslyunseendata.Therecentavailabilityofalargebodyofdatainthepublicdomain,coupledwiththespecificationofevaluationstandards,hasresultedinuniformdocumentationoftestresults,thuscontributingtogreaterreliabilityinmonitoringprogress（corpusdevelopmentactivitiesandevaluationmethodologiesaresummarizedinchapters12and13respectively）.

Finally,advancesincomputertechnologyhavealsoindirectlyinfluencedourprogress.Theavailabilityoffastcomputerswithinexpensivemassstorage