Chapter2KeywordSpottingMethodsThischapterwillreviewindetailthethreeKWS - PDF document

Chapter2KeywordSpottingMethodsThischapterwillreviewindetailthethreeKWSmethods,LVCSRKWS,AcousticKWSandPhoneticSearchKWS,followedbyadiscussionandcomparisonofthemethods.2.1LVCSR-BasedKWSPerformingKWSontextualdatabasesisrelativelystraightforward.Thetextisperusedforagivenlistofwordsandthelocationofthewordsistaggedwithinthetext.Translatingthismethodforuseinspeechdatabasesisatwo-stageprocess.First,anLVCSRengineisemployedtotransformtheentirespeechsignalintotext.TheLVCSRengineperformsthesearchforthemostprobablesequenceofwordsbasedontheViterbisearchalgorithm,usingacousticmodels,alargelexiconofwordsandalanguagemodel.Inthesecondstage,theKWSmechanismutilizesestablishedtext-basedsearchmethodstolocatethekeywordswithinthetext.Anindexingphasecanbeperformedontheresultingtextinordertoacceleratethesearchresponsetime.ThismethodwillbereferredtoasLVCSR-basedKWS.illustratesthetwosequentialstagesinvolvedinLVCSR-basedKWS.2.2AcousticKWSAnothercommonKWSmethodisAcousticKWS.Usingthismethod,theenginedoesnotattempttotranscribetheentirestreamofspeech.LiketheLVCSR-basedmethod,thismethodemploystheViterbisearch.Thatis,thesystememploysaspeechrecognitionengineonthespeech.However,ratherthanalargevocabularywhichisintendedtocoverallpotentiallyspokenwords,asmallersetofdesignatedkeywordsisusedastherecognitionvocabulary(Thambiratnam2005)andgeneralspeechmodels(aspartoftheacousticalmodels)areusedtomodelA.Moyaletal.,PhoneticSearchMethodsforLargeSpeechDatabasesSpringerBriefsinElectricalandComputerEngineering,DOI10.1007/978-1-4614-6489-1_2,SpringerScience+BusinessMediaNewYork2013 non-keywordspeech(Szokeetal.2005).Thus,acousticKWScanbeperformedinonlyonestage,asillustratedinFig.2.3PhoneticSearchKWSAsitsnamesuggests,phoneticsearchKWSutilizesaphoneticsearchengine.Intherststage,aphonemedecoderisemployedoncetotransformthespeechinputintoatextualsequence.However,ratherthanproducingastringofwords,thedecodertransformsthespeechsignalintoastring(orlattice)ofphonemes(Amiretal.2001; Second stage - Keyword DetectionKnowledge SourcesAcousticModelsLanguageModelLexiconDecoderFront-EndProcessingInputSpeechAcousticfeature vectorsKeywordsequencesKeywordDetectionKeywordListRecognizedword sequence Fig.2AnLVSCRkeywordspottingsystem–one-timetransformationofaspeechdatabase(DB)intoatextualwordDBandKWSengine ProcessingInputSpeechKnowledge SourcesAcousticfeature vectorsKeywordsequencesDecoder AcousticModelsKWPronunciation Fig.3Anacoustickeywordspottingsystem82KeywordSpottingMethods YuandSeide2004;ThambiratnamandSridharan2005).Inthesecondstage,thephoneticsearchengineemploysadistancemeasuretocomputethetextualdistancebetweenthephonemesequencesthatcorrespondtothekeywordvocabularyandthephonemesequenceswithinthephonemestring(Alon2005).AsshowninFig.,thephoneticsearchengineusestwotypesofinputdata:alistofkeywords,whereeachwordisrepresentedbyasequenceofphonemes,andaspeechdatabasewhichhasbeenrunthroughaphonemedecodertoproduceasequenceofrecognizedphonemes.2.4Discussion:WhyPhoneticSearch?EachofthethreeKWSmethodspresentedabovehasadvantagesandshortcomings.Thecrucialparameterstoevaluateareresponsetime,KWSperformance,andkeywordexibility(JamesandYoung1994;DharanipragadaandRoukos2002;Mamouetal.2007;ThambiratnamandSridharan2007;Schneider2011).2.4.1ResponseTimeIntermsofoverallcomputationalcomplexity,LVCSR-basedKWSandphoneticsearchKWSbothimplementadoublestageprocess:(1)transformationofspeechtotext(wordsequencesinthecaseofLVCSRandphonemesequencesinthecaseofphoneticsearch)and(2)akeywordsearch(word-basedinthecaseofLVCSRandphoneme-basedinthecaseofphoneticsearch).Acoustic-basedKWS,ontheotherhand,isperformedinonestageandoperatesonthespeechitselfwithnotextual AcousticModels PhonemeLMInputSpeech DBPhonemeDecoderTextual PhonemeSequence DBPhonetic SearchKWS EngineKeywordhypotheses KW List Fig.4Aphoneticsearchsystem–one-timetransformationofaspeechDBtoatextualphonemeDBandKWSphoneticsearchengine2.4Discussion:WhyPhoneticSearch?9 transformation.AlthoughakeywordsearchthatisimplementedonfullytranscribedtextintheLVCSRmethodisfast(particularlyifthetexthasalsobeenindexed),itisusuallyatadisadvantageincomparisontothephoneticsearchandacousticmethodsduetothefactthatanLVCSRenginedemandsalargevocabularyandacomplexlanguagemodeltoproducerecognitionresults,thusresultinginahighlevelofcomplexityduringthepre-processingstage.Thephoneticsearchmethodperformsphonemerecognitionusingphonemetransitionprobabilities(di-phones)withnolexiconorwordlevellanguagemodel.Duringthesearchstage,however,phoneticsearchKWSusesatextualsequencedistancemeasurethatrequiresmorecomputation.Thisisbecausethephoneticsearchmustgenerateword-levelhypothesesbasedonphonemesequences,whileinLVCSR-basedKWSthetextualoutputisalreadyword-level(Burgetetal.2006).Incontrast,theacoustic-basedKWSusesavocabularyconsistingonlyofthekeywordsanddoesnotrequirealanguagemodelatall.Becausetheacoustic-basedmethodoperatesonthespeechitselfandrequiresonlyasmallvocabulary,itisappropriateforreal-timekeywordspottingorKWSinsmallspeechdatabases.However,thismeansthatgeneralspeechmustbewell-modeled(Thambiratnam2005)toavoidextensiveoverdetection(falsealarms).2.4.2KWSPerformanceThespontaneousspeechandpoorrecordingqualityofspeechdatabasesoftenleadstodecientLVCSRperformance(Butzbergeretal.1992;Cardilloetal.2002).Thelargenumberofdisuencies,includingmispronouncedwords,falsestarts,lledpauses,overlappingspeech,speakernoisesandbackgroundnoisefoundinsponta-neousspeech(Butzbergeretal.1992;GishriandSilber-Varod2010)oftenresultsinoutputsstrewnwithwordinsertions,deletionsandsubstitutions.Thusthe“mostprobable”wordsequencesproducedbytheenginemaynotadequatelyreecttheactualinputspeech.This,inturn,affectsthereliabilityofthekeywordsearch.Thesameistruewithregardtophoneticsearchresults.Poorphonemerecogni-tionmayyieldlowerkeywordrecognitionperformanceincomparisonwiththeacousticKWSmethod,whichworksonthespeechitselfbysearchingforaspecicsequenceofphonemeswithouttextualtransformation.2.4.3KeywordFlexibilityIncomparisontothephoneticsearchmethod,whichrunsonsequencesofphonemesratherthanwords,theLVCSRmethodisatadisadvantagewhenitcomestokeywordexibility(Cardilloetal.2002;Burgetetal.2006;Wallaceetal.2007).Usingthephoneticsearchmethodallowsapplicationuserstotalfreedominchangingthedesignatedkeywords,sincethetextualtransformationintophonemes102KeywordSpottingMethods isnotrestrictedbyavocabulary.Addingnewkeywordsisasimpleprocedurethatentailsre-runningthephoneticsearchonthephonemesequences,butdoesnotrequirere-runningthephonemedecoder.ThetextualtransformationproducedbyanLVCSRengine,ontheotherhand,isconstrainedbytherecognitionvocabularyandthelanguagemodelemployed.Thus,unlessthedesignatedkeywordswerepartoftheoriginalrecognitionvocabulary,theycannotbechangedwithoutrepeatingtherecognitionprocess(Clementsetal.2001;Cardilloetal.2002;Szokeetal.2005;MamouandRamabhadran2008).Sincekeywordsareinmanycasesnamesordomain-specicvernacular,theyareoftennotfoundinstandardlexicons(Wallaceetal.2007;GishriandSilber-Varod2010).ThisisasubstantialshortcomingoftheLVCSRmethod.Acoustic-basedKWSalsorepresentsanimpracticalsolutionforsearchinglargedatabasesthatrequirerapidandexiblesearchingcapabilities.Becauseitconsistsofonlyonestage,theentireprocessneedstobere-runonthespeechdatabaseeachtimeanewkeyworddictionaryisintroduced.Themajorityofapplicationsrequirekeywordexibility,aswellastheshortestpossibleresponsetimewhensearchingverylargespeechdatabases,makingthephoneticsearchKWSmethodmoreattractivethantheLVCSRandacoustic-basedoptionswhensearchingverylargespeechdatabases.Thus,thefocusofthefollow-ingchapterswillbeonphoneticsearchKWS,andtheimplementationofanalgorithmforthereductionofcomputationalcomplexityinthephoneticsearchKWSprocess.2.4Discussion:WhyPhoneticSearch?11