Reaching over the gap外文文献.docx

Reaching over the gap外文文献.docx

《Reaching over the gap外文文献.docx》由会员分享，可在线阅读，更多相关《Reaching over the gap外文文献.docx（13页珍藏版）》请在冰豆网上搜索。

Reachingoverthegap外文文献

Reachingoverthegap:

Areviewofeffortstolinkhumanandautomaticspeechrecognitionresearch

1.Introduction

ldalgorithmsthatareabletoBoththeresearchfieldsofhumanspeechrecognition（HSR）andautomaticspeechrecognition（ASR）investigate（partsof）thespeechrecognitionprocess.Thetworesearchareasarecloselyrelatedsincetheybothstudythespeechrecognitionprocessandthecentralissueofbothiswordrecognition.However,theirresearchobjectives,theirresearchapproaches,andthewayHSRandASRsystemsdealwithdifferentaspectsofthewordrecognitionprocessdifferconsiderably（seeSection2）.Inshort,inHSRresearch,thegoalistounderstandhowwe,aslisteners,recognisespokenutterances.ThisisoftendonebybuildingcomputationalmodelsofHSR,whichcanbeusedforthesimulationandexplanationofbehaviouraldatarelatedtothehumanspeechrecognitionprocess.TheaimofASRresearchistobuildalgorithmsthatareabletorecognisethewordsinaspeechutteranceautomatically,underavarietyofconditions,withtheleastpossiblenumberofrecognitionerrors.MuchresearcheffortinASRhasthereforebeenputintotheimprovementof,amongstothers,signalrepresentations,searchmethods,andtherobustnessofASRsystemsinadverseconditions.

Onemightexpectthatinthepastthiscommongoalwouldhaveresultedinclosecollaborationsbetweenthetwodisciplines,butinreality,theoppositeistrue.ThislackofcommunicationismostlikelytobeattributedtoanotherdifferencebetweenASRandHSR.AlthoughbothASR

andHSRclaimtoinvestigatethewholerecognitionprocessfromtheacousticsignaltotherecognisedunits,anautomaticspeechrecognisernecessarilyisanend-toendsystem–itmustbeabletorecognisewordsfromtheacousticsignal–whilemostmodelsofHSRonlycoverpartsofthehumanspeechrecognitionprocess（Nearey,2001;MooreandCutler,2001）.Furthermore,inASR,thealgorithmsandthewaytotraintheASRsystemsarecompletelyunderstoodfromamathematicalpointofview,

butinpracticeithassofarprovedimpossibletogetthedetailssufficientlyrighttoachievearecognitionperformancethatisevenclosetohumanperformance.Humanlisteners,ontheotherhand,achievesuperiorperformance,butmanyofthedetailsoftheinternalprocessesare

unknown.

Despitethisgapthatseparatesthetworesearchfields,thereisagrowinginterestinpossiblecross-fertilisation（Furui,2001;Hermansky,2001;Huckvale,1998;KirchhoffandSchimmel,2005;Moore,1995;MooreandCutler,2001;Pols,1999;Scharenborg,2005a,b;Scharenborgetal.,2005;tenBosch,2001）.（Forahistoricalbackgroundontheemergenceofthisgap,seeHuckvale（1998）.）Thisis,forinstance,clearlyillustratedbytheorganisationoftheworkshoponSpeechrecognitionaspatternclassification（July11–13,2001,Nijmegen,TheNetherlands）,theorganisationofthespecialsessionBridgingthegapbetweenhumanandautomaticspeechprocessingatInterspeech2005（September6,2005,Lisbon,Portugal）andofcoursewiththecomingaboutofthis‘SpeechCommunication’specialissueonBridgingthegapbetweenhumanandautomaticspeechprocessing.

ItisgenerallyacknowledgedwithintheASRcommunitythattheimprovementinASRperformanceobservedinthelastfewyearscantoalargeextentbeattributedtoanincreaseincomputingpowerandtheavailabilityofmorespeechmaterialtotraintheASRsystems（e.g.,Bourlardetal.,1996;MooreandCutler,2001）.However,theincremental

performanceisasymptotingtoalevelthatfallsshortofhumanperformance.ItistobeexpectedthatfurtherincreasingtheamountoftrainingdataforASRsystemswillnotresultinrecognitionperformancesthatareevenapproachingthelevelofhumanperformance（Moore,2001,2003;MooreandCutler,2001）.Whatseemstobeneededisachangeinapproach–evenifthismeansaninitialworseningoftherecognitionperformance（Bourlardetal.,1996）.AspointedoutbyMooreandCutler（2001）:

‘‘trueASRprogressisnotonlydependentontheanalysisofevermoredata,butonthedevelopmentofmorestructuredmodelswhichbetterexploittheinformationavailableinexistingdata’’.ASRengineershopetogetcluesaboutthose‘‘structuredmodels’’fromtheresultsofresearchin

HSR.Thus,fromthepointofviewofASR,thereishopeofimprovingASRperformancebyincorporatingessentialknowledgeaboutHSRintocurrentASRsystems（Carpenter,1999;DusanandRabiner,2005;Furui,2001;Hermansky,1998;MaierandMoore,2005;Moore,2003;MooreandCutler,2001;Pols,1999;Scharenborgetal.,2007;Strik,2003,2006;Wright,2006）.

WithrespecttothefieldofHSR,specificstrandsinHSRresearchhopetodeployASRapproachestointegratepartialmodulesintoaconvincingend-to-endmodel（Nearey,2001）.Aspointedoutabove,computationalmodelsofHSRonlymodelpartsofthehumanspeechrecognitionprocess;anintegralmodelcoveringallstagesofthehuman

speechrecognitionprocessdoesnotyetexist.Themostconspicuouspartoftherecognitionprocessthat,untilrecently,virtuallyallmodelsofhumanspeechrecognitiontookforgrantedisamodulethatconvertstheacousticsignalintosomekindofsymbolicsegmentalrepresentation.

So,unlikeASRsystems,mostexistingHSRmodelscannotrecogniserealspeech,becausetheydonottaketheacousticsignalastheirstartingpoint.In2001,NeareypointedoutthattheonlyworkingmodelsoflexicalaccessthattakeanacousticsignalasinputwereASRsystems.Mainstream

ASRsystemshoweverareusuallyimplementationsofaspecificcomputationalparadigmandtheirrepresentationsandprocessesneednotbepsychologicallyplausible.IntheircomputationalanalysisoftheHSRandASRspeechrecognitionprocess,Scharenborgetal.（2005）showedthat

someASRalgorithmsservethesamefunctionsasanalogousHSRmechanisms.Thusdespitefirstappearances,thismakesitpossibletousecertainASRalgorithmsandtechniquesinordertobuildandtestmorecompletecomputationalmodelsofHSR（RoyandPentland,2002;Scharenborg,2005b;Scharenborgetal.,2003,2005;Wadeetal.,2002;Yuetal.,2005）.

Furthermore,withinthefieldofASRtherearemany（automaticallyorhand-annotated）speechandlanguagecorpora.ThesecorporacaneasilyandquicklybeanalysedusingASRsystems,sinceASRsystemsandtoolsareabletoprocesslargeamountsofdataina（relatively）short

time.ThismakesASRtechniquesvaluabletoolsfortheanalysisandselectionofspeechsamplesforbehaviouralexperimentsandthemodellingofhumanspeechrecognition（deBoerandKuhl,2003;KirchhoffandSchimmel,2005;Pols,1999）.

ResearchersfrombothASRandHSRarethusrealisingthepotentialbenefitoflookingattheresearchfieldontheothersideofthegap.Thispaperintendstogiveacomprehensiveoverviewofpastandpresenteffortstolinkhumanandautomaticspeechrecognitionresearch.ThefocusofthepaperisonthemutualbenefitstobederivedfromestablishingclosercollaborationsandknowledgeinterchangebetweenASRandHSR.FirstabriefoverviewofthegoalsandresearchapproachesinHSRandASRisgiveninSection2.InSection3,wediscusstheperformance

differencebetweenmachinesandhumanlistenersforvariousrecognitiontasksandwhatcanbelearnedfromcomparingthoserecognitionperformances.Section4discussesapproachesforimprovingthecomputationalmodellingofHSRbyusingASRtechniques.Section5presents

anoverviewoftheresearcheffortsaimedatusingknowledgefromHSRtoimproveASRrecognitionperformance.Thepaperendswithadiscussionandconcludingremarks.

2.Humanandautomaticspeechrecognition

Inthissection,theresearchfieldsandapproachesofhumanspeechrecognition（Section2.1）andautomaticspeechrecognition（Section2.2）willbediscussedbriefly.Amoredetailedexplanationofthetworesearchfieldswouldbebeyondthescopeofthisarticle;thereaderisreferredtotextbookaccountssuchasHarley（2001）foranin-depthcoverageoftheresearchfieldofhumanspeechrecognition,andtotextbookaccountssuchasRabinerandJuang（1993）orHolmesandHolmes（2002）,foran

explanationoftheprinciplesofautomaticspeechrecognition.Comprehensivecomparisonsoftheresearchgoalsandapproachesofthetwofields（Huckvale,1998;MooreandCutler,2001;Scharenborgetal.,2005）,aswellascomparisonsofthecomputationalfunctioningandarchitectures

（Scharenborgetal.,2005）ofthespeechrecognitionprocessinautomaticrecognitionsystemsandhumanlistenerscanalsobefoundintheliterature.Furthermore,DusanandRabiner（2005）provideanextensivecomparisonbetweenhumanandautomaticspeechrecognitionalongsixkeydimensions（includingthearchitectureofthespeechrecognitionsystem）ofASR.

2.1.Humanspeechrecognition

Toinvestigatethepropertiesunderlyingthehumanspeechrecognitionprocess,HSRexperimentswithhumansubjectsareusuallycarriedoutinalaboratoryenvironment.Subjectsareaskedtocarryoutvarioustasks,suchas:

•Auditorylexicaldecision:

Spokenwordsandnon-wordsarepresentedinrandomordertoalistener,whoisaskedtoidentifythepresenteditemsasawordoranon-word.

•Phoneticcategorisation:

Identificationofunambiguousandambiguousspeechsoundsonacontinuumbetweentwophonemes.

•Sequencemonitoring:

Detectionofatargetsequence（largerthanaphoneme,smallerthanaword）,whichmaybeembeddedinasentenceorlistofwords/nonwords,orinasinglewordornon-word.

•Gating:

Awordispresentedinsegmentsofincreasingdurationandsubjectsareaskedtoidentifythewordbeingpresentedandtogiveaconfidenceratingaftereachsegment.

Intheseexperiments,va