Traditional concatenative speech synthesis systems use a number of heuristics to define the target and concatenation costs, essential for the design of the unit selection component. In contrast to these approaches, we introduce a general statistical modeli
ofthetargetandconcatenationcostsintraditionalunitselection.
Thestatisticalframeworkjustoutlinedissimi-lartotheoneusedinspeechrecognition(Jelinek,1976).Wealsouseseveraltechniquesthathavebeenverysuccessfullyappliedtospeechrecogni-tion.Forinstance,inthispaper,weshowhow logP(u)(theconcatenationcost)canbeaccu-ratelyestimatedusingastatisticaln-gramlanguagemodeloverunits.Twoquestionsnaturallyarise.(a)Howcanwecollectatrainingcorpusforbuild-ingastatisticalmodel?Ideally,thetrainingcor-puscouldbehuman-labeled,asinspeechrecog-nitionandothernaturallanguageprocessingtasks.Butthisseemedimpracticalgiventhesizeoftheunitinventory,thenumberofutterancesneededforgoodstatisticalestimates,andourlimitedresources.Instead,wechosetouseatrainingcorpusgener-atedbyanexistinghigh-qualityunitselectionsys-tem,thatoftheAT&TNaturalVoicesProduct.Ofcourse,buildingastatisticalmodelonthatoutputcan,atbest,onlymatchthequalityoftheorigi-nal.But,itcanserveasanexploratorytrialtomea-surethequalityofourstatisticalmodeling.Aswewillsee,itcanalsoresultinasynthesissystemthatissigni cantlyfasterandmodularthantheoriginalsincetherearewell-establishedalgorithmsforrep-resentingandoptimizingstatisticalmodelsofthetypewewillemploy.Tofurthersimplifytheprob-lem,wewillusetheexistingtraditionaltargetcosts,providingstatisticalestimatesonlyoftheconcate-nationcosts( logP(u)).
(b)Whatarethebene tsofastatisticalmodelingapproach?
(1)High-qualitycostfunctions.Oneissuewithtraditionalunitselectionsystemsisthattheircostfunctionsaretheresultofthefollowingcompromise:theyneedtobecomplexenoughtohaveaperceptualmeaningbutsimpleenoughtobecomputedef ciently.Withourstatisticalmodelingapproach,thelabelingphasecouldbeperformedof inebyahighlyaccurateunitselec-tionsystem,potentiallyslowandcomplex,whiletherun-timestatisticalsystemcouldstillbefast.Moreover,ifwehadaudioavailableforourtrainingcorpus,wecouldexploitthatintheinitiallabel-ingphaseforthedesignoftheunitselectionsystem.(2)Weighted nite-statetransducerrepresenta-tion.Inadditiontothealreadymentionedsynthesisspeedandtheopportunityofhigh-qualitymeasuresintheinitialof inelabelingphase,anotherbene tofthisapproachisthatitleadstoanaturalrepresen-
tationbyweightedtransducers,andhenceenablesustobuildaunitselectionsystemusinggeneraland exiblerepresentationsandmethodsalreadyinuseforspeechrecognition,e.g.,thosefoundintheFSM(Mohrietal.,2000),GRM(Allauzenetal.,2004)andDCD(Allauzenetal.,2003)libraries.Otherunitselectionsystemsbasedonweightedtransducerswerealsoproposedin(Yietal.,2000;BulykoandOstendorf,2001).
(3)Unitselectionalgorithmsandspeed-up.Wepresentanewunitselectionsystembasedonsta-tisticalmodeling.Weusedweightedautomataandtransducersfortherepresentationofthecompo-nentsofthesystemanddesignedanewandef cientcompositionalgorithmmakinguseofstringpoten-tialsfortheircombination.Theresultingstatisticalunitselectionisshowntobeabout2.6timesfasterthanthelastreleaseoftheAT&TNaturalVoicesProductwhilepreservingthesamequality,andof-fersmuch exibilityfortheuseandintegrationofnewandmorecomplexcomponents.
2UnitSelectionMethods
2.1
OverviewofaTraditionalUnitSelectionSystem
ThissectiondescribesindetailthecostfunctionsusedintheAT&TNaturalVoicesProductthatwewilluseasthebaselineinourexperimentalresults,see(Beutnageletal.,1999a)formoredetailsaboutthissystem.Inthissystem,unitselectionisbasedon(HuntandBlack,1996)butusingunitscorre-spondingtohalfphonesinsteadofphones.LetUbethesetofrecordedunits.Twocostfunctionsarede ned:thetargetcostCt(fi,ui)isusedtoestimatethemismatchbetweenthefeaturesofthefeaturevectorfiandtheunitui;theconcatena-tioncostCc(ui,uj)isusedtoestimatethesmooth-nessoftheacousticsignalwhenconcatenatingtheunitsuianduj.Givenasequencef=fselectioncanthenbe1...fnoffeaturevectors,unitformu-latedastheproblemof ndingthesequenceofunitsu=u1...unthatminimizesthesetwocosts:
u=argminu∈Un
(
nnCt(fi,ui)+ Cc(ui 1,ui))i=1
i=2
Inpractice,notallunitsequencesofagivenlength
areconsidered.Apreselectionmethodsuchastheoneproposedby(Conkieetal.,2000)isused.Thecomputationofthetargetcostcanbesplitintwoparts:thecontextcostCpthatisthecomponentofthetargetcostcorrespondingtothephoneticcon-text,andthefeaturecostCfthatcorrespondsthe
(2004). Statistical modeling for unit selection in speech synthesis, In 42nd Meeting Assoc. Computational Linguistics (ACL 2004), Barcelona, Spain. Allouche...
Statistical Modeling for... 8页 免费 Towards ...for unit selection in speech synthesis,” in ...criterion,” in Autumn Meeting of ASJ, 2003, ...
“Creating a Database of Speech in Noise for Unit Selection Synthesis”, ...“Improving the Understandability of Speech Synthesis by Modeling Speech in ...
? ? Statistical parametric speech generation and synthesis HMM-based speech ...? ? ? ? Concatenative synthesis Sample-based (unit selection) High quality...
Prosodic modeling in t... 暂无评价 10页 免费 Homograph disambiguati... ...(1996) Unit selection in a concatenative speech synthesis system using a ...
Thus, many di?erent statistical methods used for...So, the feature selection in this paper is ... for high quality Mandarin speech synthesis. ...
A UNIT SELECTION APPR... 暂无评价 6页 免费...statistical taggers [Petasis et al., 1999], ...From the speech synthesis point of view, input ...
A UNIT SELECTION APPR... 暂无评价 6页 免费...statistical taggers [Petasis et al., 1999], ...From the speech synthesis point of view, input ...
热门文档