Text-based and Signal-based Prediction of Break Indices and Pause Durations

  • 格式:pdf
  • 大小:194.75 KB
  • 文档页数:4

Text-basedandSignal-basedPredictionofBreakIndicesandPauseDurations

HartmutR.Pfitzinger&UweD.Reichel

InstituteofPhoneticsandSpeechCommunication

UniversityofMunich,Germany

{hpt;reichelu}@phonetik.uni-muenchen.de

Abstract

Therelationbetweensymbolicandsignalfeaturesofprosodic

boundariesisexperimentallystudiedusingpredictionmethods.

Text-basedbreakindexpredictionturnsouttobefairlygood,

butsignal-basedpredictionandpausedurationpredictionper-formworse.Apossiblereasonisthatrandomsignalfeature

variations,asusuallyproducedbyhumans,arehardtopredict.

1.Introduction

Speakersdividetheirutterancesintoprosodicphrasesseparated

byweakorstrongprosodicboundaries.Theseboundariescan

beperceptuallyclassifiedaccordingtotheToBIbreakindices,

ortheycanbedescribedbymeasurementsofprosodicfeatures

ofthespeechsignal,i.e.thepausedurationandthelocalcon-toursofF0,speechrate,amplitude,andvoicequalityinthe

vicinityoftheprosodicboundary.Thegoalofthepresentstudy

istoinvestigatetherelationshipbetweensymbolicandsignal

representationsofprosodicboundariesbymeansofthreepre-dictionmethods(seeFig.1).

Prosodicphrasebreakpredictioncanbedecomposedinto

twotasks:breaklocalizationandpredictionofitsphoneticreal-

ization.Forthelatterthisstudyconcentratesonpauseduration.Rule-basedorstatisticalapproachestacklebothtasksbyex-

ploitingpartofspeech(POS),syntactical,andrhythmicalinfor-

mation.Forrule-basedbreaklocalizationLiberman&Church

[9]usethefact,thatcertainPOSclassesoccurpreferablyphraseinitially(chinks)whileothersdonot(chunks).Gee&Grosjean

[7]andBachenko&Fitzpatrick[2]imposesyntacticallymoti-

vatedperformancestructuresonanutterance.Thedistanceof

twoadjacentwordsinthesestructuresgivenintheleveloftheircommonnodedeterminesthebreakstrengthandthusthepause

durationbetweenthesewords.Therule-basedKeller-Zellner

algorithm[19]alsoincorporatesrhythmicalconstraints.

InsomestatisticalapproachesphrasebreaksarepredictedfromagivenpartofspeechsequencebyamodifiedMarkov

Tagger.Forexample,Black&Taylor’stagger[3]isbased

onconditionalemissionprobabilitiesforPOSsequencesco-

occurringwiththebreaktypegivenatthecorrespondingstateandconditionaltransitionprobabilitiesofthisbreaktypefora

givenbreaktypehistory.Others(e.g.[1])useclassificatorsas

CART[4]inordertopredictphrasebreaksandpauselengths

fromasetoflinguisticfeaturesaspartofspeechandwordandsyllabledistances.CARTsarewellsuitedforthistaskdueto

theirabilitytocopewithcategoricalandcontinuoustypesof

dependentandindependentvariables.

In2002,Mixdorff[11]investigated,amongotherprosodic

properties,pausedurationsintheIMSRadioNewsCorpus[15].Meanvaluesandstandarddeviationsofpausesatsen-

tenceboundarieswere716msand336ms,respectively,andof

pauseswithinsentenceswere327msand132ms,respectively.2.Method

Weinvestigatedthepredictabilityofbreakindicesandpausedu-

rationsbytext-basedandbysignal-basedfeatures(seeFig.1).

Inordertoaccountforhumandurationperception,pausedu-

rationswererepresentedonalogarithmicscale.ForbreakindexpredictionweusedC4.5decisiontrees[14],forpause

durationsCARTs(classificationandregressiontrees[4])pro-

videdbyR.Additionally,weanalysedthecorrelationbetween

prosodicmeasuresandpauseduration.

2.1.Data

ThepresentstudyisbasedontheIMSRadioNewsCorpus[15]whichconsistsofGermannewstextsreadbyprofessional

speakers.Originally,thedataisautomaticallysegmentedinto

phonemesaccordingtotheGermanSAM-PAinventoryfol-

lowedbysomemanualrefinements.ProsodyofthedatawasmanuallylabelledfollowingtheGToBIconventions[10].

Priortoanalysisofthedataasubstantialcorrectionwas

necessary.Weomittednewsrepetitionsandmanuallycorrected

thephoneboundariesinordertoachievereliablepausedura-tionandPLSRmeasures(perceptuallocalspeechrate[12,13]).

Particularly,wereplacedcanonicaltranscriptionsbytheiractual

phoneticrealizations,insertedglottalstops,whichhaveunfortu-

natelybeenabsent,andreplaced/Vowel/-/R/-pairsbytheircor-responding/Vowel/-/6/-diphthongs,especially/@/-/R/by/6/.

Finally,ourdatacomprises28newsarticlesreadbyone

malespeaker.Itconsistsof16285phones,2384wordtokens,

and970wordtypes,whichisapproximatelyhalfoftheorigi-naldata.Fortext-basedpredictionallfinalwordsofeachsig-

nalwereexcluded,becausesubsequentpausedurationnaturally

couldnotbemeasured.80%ofthisdatawereusedfortraining

andtheremainderastestset.Thedataforn-gramandcollo-cationmodelingcomesfromdiversewrittennewscorporaand

consistsof327821wordtokensand42741types.

CARTCART(Part of Speech, Position, Cohesion)

C4.5

S I G N A L SBreak Indices

Pause Durations

Lin. RegressionC4.5S Y M B O L S

(Prosodic Contours of Intonation, Speech Rate, Amplitude)

Figure1:Text-based(top-down)andsignal-based(bottom-up)

predictionsofbreakindicesandpausedurationsinthisstudy.CategoricaldataarepredictedbyC4.5decisiontrees,andcon-

tinuousdatabyCARTsorbylinearregression.