NORTHWESTERNUNIVERSITY
TheVoiceofExperience:CausalInferenceinPhonotacticAdaptation
ADISSERTATION
SUBMITTEDTOTHEGRADUATESCHOOLINPARTIALFULFILLMENTOFTHE
REQUIREMENTS
forthedegree
DOCTOROFPHILOSOPHY
FieldofLinguistics
ByThomasDenby
EVANSTON,ILLINOIS
June2019
2
Abstract
Successfullygrapplingwiththewidespreadlinguisticvariationofdailyliferequires
speakerstoadapttosystematicvariationintheenvironmentwhilediscardingincidental
variation,basedontheirpriorexperience.Inthecaseofphonotactics,speakers’prior
experienceisthattalkerswhodifferintheirlanguagebackgroundarelikelytovaryintheir
phonotacticgrammars,whiletalkerswhosharealanguagevarietyareunlikelytodoso.As
such,wepredictthatwhenspeakersareexposedtomultipletalkerswhosephonotactics
vary,andthosetalkersdifferintheirlanguagebackground,listenerswillinferthevariation
issystematicandadapt.Conversely,ifthetalkerssharealanguagebackground,listeners
willinferthevariationisincidental,andnotadapt.
InStudy1,wetestedthispredictioninaperceptionexperiment,byexposing
listenerstotwotalkers,eachofwhomexhibitedadifferentphonotacticconstraint,ina
recognitionmemorytask.InExperiment1,whenlistenerswereexposedtotalkerswho
differedintheirlanguagebackground(1Englishvs.1Frenchtalker),theyshowedahigh
degreeofadaptation;whenthetalkerssharedalanguagebackground(2Englishor2
Frenchtalkers),listenersshowedalow-to-moderatedegreeofadaptation.InExperiment2,
weexaminedthegranularityoflistenerknowledgeofvariationinnon-nativephonotactics
byincludinganovelconditionwithtwonon-nativetalkers(1Hindivs.1Hungariantalker).
Listenersshowedahighdegreeofadaptationevenwhenbothtalkerswerenon-native
speakerswithdifferentlanguagebackgrounds,suggestingthatlistenersmakedistinctions
betweendifferentnon-nativelanguagephonotactics.
3
InStudy2,weexaminetheroleofcausalinferenceinspeechproduction.Recent
worksuggestsadaptationinproductionmaydifferfromperception,asproductionmay
utilizesimpleassociativelearningmechanismsthatmaynottakehigh-levelindexical
featuresintoaccount.Weexplorethisquestionusingamodifiedtonguetwisterparadigm,
inwhichparticipantsrepeatsyllablesequencesfromtwomodeltalkers,witheachtalker
exhibitingadifferentphonotacticconstraint.MirroringStudy1,modeltalkerseither
sharedanon-nativelanguagebackground;sharedanativelanguagebackground;or
differedintheirlanguagebackground(thelanguagesbackgroundsinquestionwere
GermanandEnglish).Inaddition,acontrolconditionwasincludedfollowingprevious
tonguetwisterexperiments,inwhichthephonotacticconstraintwasconditionedonthe
identityoftheadjacentvowel.Resultswerelargelyinconclusive—therewassomeevidence
ofincreasedadaptationwhenparticipantswereexposedtomodeltalkerswithdifferent
languagebackgrounds,buttheeffectwasinconsistent.Inaddition,noeffectwasfoundin
thecontrolcondition.
Together,theseresultssuggestthatphonotacticadaptationisflexible,but
constrainedbythecausalinferenceslistenersdrawfromtheirpriorexperience,
particularlyinperception.
4
Acknowledgements
It’shardtolookbackoverthelastsixyearsandadequatelyputintowords,oreven
fullyaccountfor,theprofoundhelpmyfriends,family,andcolleagueshavegivenmetoget
tothispoint.I’lldomybest.
Firstandforemost,I’dliketothankmyadvisorandmentor,MattGoldrick.It’sbeen
anhonortoworkwithyou,andIfeltsoprofoundlyluckysoofteninmygraduatecareerto
haveyourexpertguidance,calmdemeanor,andinfinitewisdominmycorner.You’rea
modelofwhatagoodmentor,teacher,andresearchershouldbe,andIfeelsothankfulto
havehadthatmodelthroughoutgradschool.I’llmissourmeetings,spentuntangling
knottyideasandmakingobscureSimpsonsreferences.Iwouldnotbetheresearcherand
thinkerIamwithoutyourhelp,norwouldthisdissertationhavebeenpossible.Thankyou
foreverything.(AndsorryaboutallthosedeadlinesImissed…)
ThankyouAnnBradlowandJenniferCole,mycommitteemembers,foryour
expertiseandcriticaleye.Youeachconsistentlychallengedmyassumptionsinawaythat
mademethinkmoredeeplyaboutmywork.Yourexperienceandacumenmadethe
dissertationstronger,andyourencouragementandpositivitybuoyedmyspiritsalongthe
way.
ThankyoutotheNationalScienceFoundation,fortheirgenerousfundingand
support(grant#1728173).
ThankyoutothemembersofSoundLab,pastandpresent,whoheardmetalkabout
thisworkinitsvariousformsfor5yearsandprovidedinvaluablefeedbackalongtheway
5
(andlistenedtomecomplainabouthand-codingspeecherrorseveryweek).Specialthanks
tolabmembersEmilyCibelli,ErinGustafson,NicoleMirea,andMariaGavinoinparticular
foryouradvice,support,andfriendship.ThankyoutoallofmyfellowLinggradstudents
overtheyears,fortheircamaraderieandsupport.
ThankyoutoChunChan,whosetechnicalwizardryandabilitytosolvedifficult
logisticalproblemsmadehugeportionsofthisprojectpossible.Thankyoualsotothemany
brilliantlinguistswhogavemeadviceandideasfornovelconditionsalongtheway,
especiallyMelissaBaese-BerkandGaryDell.
ThankyoutoKatHall,mycohort-mate,confidant,andbestbud.Takingthisjourney
withyouhasbeenoneofthegreatjoysofmylife.Youweremyrockthroughoutgraduate
school,andhangingoutwithyouneverfailedtofillmewithjoyandappreciationforwhata
strong,brilliant,andsillypersonyouare.ToSean,thankyouforyourcompanionship,your
warmth,andyourendlessgoodness.YouandKat(andAbbeyandMiles!)aremyChicago
family,andIloveyou.
Thankyoutomyothercohort-mateNayounKim,forbeingtheperfectofficebuddy
andalwayssupportingmeinthetoughesttimes.I’llalwaysthinkbackonourgiggly,
punch-drunkdaysinSwift309spentagonizingoverourdissertationswithfondness.
ThankyoutomymanyChicagofriendsforkeepingmesane,healthy,andhappy
overthelastsixyears.ToLaura,John,andTaylor,forbeingmyfavoritestdrinkingbuddies
andgenerallythemostrefreshing,comforting,wonderfulhumanbeingsinChicago.To
Libby,forbeingwithmethroughalltheupsanddowns(#sadnesstwinsforever).Tothe
6
foundingmembersofTatLPD(EmilyCibelli,AngelaCooper,ErinGustafson,CindyBlanco,
JulieMatsubara),I’llmissourjamsessions.
ThankyoutoBriannaKaufmanandAlexKapelman,mylong-distancebestbros,who
supportedmeandletmeblowoffsteamcountlesstimesonthephone.Youarethetruest
friendsIcouldeveraskfor.
Thankyoutoallfourofmyparentalunits.ToMomandDad,mywordswilldono
justicetowhatyouhavegivenme,butheregoes.Thankyouforalwaysbeingthereforme,
alwayssupportingmeandmyeducation,alwayslendingmeanempatheticear.Thankyou
forstokingmyintellectualcuriosity,forteachingmehowtowriteandthinkcritically,for
pushingmetodomore,andforpickingmeupwhenIfall.Mostofall,thankyoufor
kindness,andforyourlove.ToJanetandSusan,thankyouforbeingthemostwonderful,
supportive,andaffectionatestepparentsIcouldeveraskfor.Iamfilledwithgratitudethat
youbothcameintomylife,andIknowI’msmarter,kinder,andmoreopenbecauseofit.
Finally,thankyoutomypartnerCassandraRose,whohasbeenbymysideforevery
upanddownofthisproject(andthereweremany).Youspentendlesshourssittingonthe
couchwithmewhilewebothclackedawayonourlaptops,supportingmywork,listening
tomekvetch,andboostingmyspiritswithfrequentboutsofextremesilliness.Your
patienceandyourgoodnessandyourkindnessareunlimited.YouarethebestpersonI
know,andIamsoluckytohaveyouinmylife.IloveyouandIlikeyou.
7
TableofContents
ABSTRACT.................................................................................................................................................2
ACKNOWLEDGEMENTS........................................................................................................................4
TABLEOFCONTENTS............................................................................................................................7
LISTOFFIGURES..................................................................................................................................11
LISTOFTABLES...................................................................................................................................13
1. INTRODUCTION...........................................................................................................................15
2. STUDY1..........................................................................................................................................22
2.1. INTRODUCTION...........................................................................................................................................................22
2.2. BACKGROUND.............................................................................................................................................................27
2.2.1. Phonotactics........................................................................................................................................27
2.2.2. AdaptationandVariation.............................................................................................................29
2.2.3. PhonotacticAdaptationandVariation...................................................................................31
2.3. EXPERIMENT1A........................................................................................................................................................33
2.3.1. Participants.........................................................................................................................................37
2.3.2. Stimuli....................................................................................................................................................38
2.3.3. Procedure.............................................................................................................................................39
2.3.4. Design.....................................................................................................................................................40
2.3.5. DataAnalysis......................................................................................................................................41
2.3.6. Results....................................................................................................................................................43
8
2.3.7. Experiment1ADiscussion.............................................................................................................45
2.4. EXPERIMENT1B........................................................................................................................................................48
2.4.1. Participants.........................................................................................................................................49
2.4.2. Stimuli....................................................................................................................................................49
2.4.3. DataAnalysis......................................................................................................................................50
2.4.4. Results....................................................................................................................................................51
2.4.5. Discussion.............................................................................................................................................52
2.4.6. Experiment1conclusion................................................................................................................54
2.5. EXPERIMENT2...........................................................................................................................................................55
2.5.1. Participants.........................................................................................................................................57
2.5.2. Stimuli....................................................................................................................................................57
2.5.3. Procedure.............................................................................................................................................59
2.5.4. Design.....................................................................................................................................................59
2.5.5. DataAnalysis......................................................................................................................................60
2.5.6. Results....................................................................................................................................................62
2.5.7. Discussion.............................................................................................................................................63
2.6. LISTENERLANGUAGEBACKGROUNDANALYSIS...................................................................................................65
2.7. GENERALDISCUSSION...............................................................................................................................................68
2.7.1. PhonotacticsandL2Acquisition................................................................................................70
2.7.2. AccentDetection................................................................................................................................72
2.8. CONCLUSION...............................................................................................................................................................73
3. STUDY2..........................................................................................................................................74
3.1. INTRODUCTION...........................................................................................................................................................74
9
3.2. BACKGROUND.............................................................................................................................................................77
3.2.1. Phonotacticlearninginspeechproduction..........................................................................77
3.2.2. Causalinferenceinphonotacticadaptationinproduction............................................79
3.2.3. CurrentStudy......................................................................................................................................81
3.3. METHODS....................................................................................................................................................................84
3.3.1. Participants.........................................................................................................................................84
3.3.2. Materials...............................................................................................................................................85
3.3.3. Procedure.............................................................................................................................................88
3.3.4. Analysis..................................................................................................................................................89
3.4. RESULTS......................................................................................................................................................................90
3.4.1. Discussion...........................................................................................................................................102
3.4.2. Production-perceptiondynamicsandphonotacticadaptation.................................107
3.5. CONCLUSION............................................................................................................................................................108
4. CONCLUSION..............................................................................................................................110
4.1. STUDY1....................................................................................................................................................................111
4.2. STUDY2....................................................................................................................................................................115
4.3. FUTUREDIRECTIONS..............................................................................................................................................117
4.4. CONCLUSIONS..........................................................................................................................................................122
5. APPENDIX...................................................................................................................................123
5.1. APPENDIXA–STUDY1VOWELACOUSTICSANALYSIS....................................................................................123
5.2. APPENDIXB–STUDY1PILOTSTUDY.................................................................................................................125
5.3. APPENDIXC–STUDY1POWERANALYSIS.........................................................................................................127
5.4. APPENDIXD–STUDY1PASSINGRATES............................................................................................................128
10
5.5. APPENDIXE–STUDY1MODELRESULTS...........................................................................................................134
5.6. APPENDIXF–TONGUETWISTERSAMPLESIZEANALYSIS..............................................................................136
5.7. APPENDIXG–STUDY2MODELRESULTS...........................................................................................................138
5.7.1. Within-conditionmodels;allsessions;[m]-[n]swapsincluded..................................138
5.7.2. Within-conditionmodels;[m]-[n]swapsexcluded..........................................................139
5.7.3. Within-conditionmodels;session1excluded;[m]-[n]swapsexcluded..................140
6. REFERENCES..............................................................................................................................142
11
ListofFigures
Figure2.1.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment1A.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment1A.Inbothpanels,errorbarsreflectbootstrapped95%confidenceintervals...........................................................................................................................................................44
Figure2.2.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment1B.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment1B.Inbothpanels,errorbarsreflectbootstrapped95%confidenceintervals...........................................................................................................................................................51
Figure2.3.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment2.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment2.Inbothpanels,errorbarsreflectbootstrapped95%confidenceinterval..............................................................................................................................................................................62
Figure3.1.Overallerrorratesbyparticipant.Colorsreflectexperimentalcondition...........92Figure3.2.Percentageoferrorsthatmaintaintheirsyllablepositionforconstrainedvs.
unconstrainedconsonants,brokendownbysessionandcondition.Alldatavisualized,including[m]-[n]swaps.Errorbarsreflect95%confidenceintervaloverparticipant;however,notethatparticipantscontributeddifferentnumbersoferrorstoeachbar.96
Figure3.3.Percentageoferrorsthatmaintaintheirsyllablepositionforconstrainedvs.unconstrainedconsonants,brokendownbysessionandcondition.[m]-[n]swapswereexcluded.Errorbarsreflect95%confidenceintervaloverparticipant;however,notethatparticipantscontributeddifferentnumbersoferrorstoeachbar.IntheVowelcondition,oneparticipantmadethemajorityoftheerrorsbuthadamuchlowermeanthanotherparticipants,resultinginaCIthatdoesnotoverlapwiththemeanforSession1,unconstrainederrors............................................................................................................99
Figure5.1.ScatterplotoffemaleFrenchtalkers’firstandsecondvowelformants.Eachpointisavowel,withvowelidentityindicatedbytheappropriateIPAsymbol.Colorindicatestalkerdifferences...................................................................................................................124
Figure5.2.Scatterplotofhitrate(%yesonfamiliaritems)ingeneralizationphasebylegalityadvantage(falsealarmrateforlegalitemsminusfalsealarmrateonillegalitems)forallexperimentsinStudy1.Eachdotrepresentsasingleparticipant;colorsrepresentwhetherparticipantspassedorfailedcriteria.LinesrepresentLoessregression;shadingrepresents95%confidenceinterval.......................................................130
Figure5.3.LegalityadvantageforExperiment1A,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval......................................................132
Figure5.4.LegalityadvantageforExperiment1B,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval......................................................132
12
Figure5.5.LegalityadvantageforExperiment2,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval......................................................133
13
ListofTables
Table2.1.SummaryofconditionsinExperiment1A,alongwithexperimentalspeakerlanguagebackground,gender,andstimulusvowels.NotethatvirtuallyalllistenerswerenativeEnglishspeakers.................................................................................................................35
Table2.2.SummaryofconditionsinExperiment2,alongwithexperimentalspeakerlanguagebackground,gender,andstimulusvowels.NotethatvirtuallyalllistenerswerenativeEnglishspeakers.................................................................................................................56
Table3.1.Summaryofconditions,alongwithmodeltalkerlanguagebackground,vowels,gender,anddegreeofadaptation.........................................................................................................84
Table3.2.Numberoferrorsanderrorratesbycondition..................................................................92Table3.3.Target/errormatrixforallconsonantsandconditions.Targetsarecolumnsand
errorsarerows.TheNoTargetcolumnreferstoerrorsthatweremadeonextrasyllables(i.e.,whenaparticipantproducedmorethan4syllablesinasingletwister).Gradientcolor-codingreflectsthenumberoferrorsforagiventarget/errorcombination...................................................................................................................................................94
Table3.4.Betweenconditioncomparison;allsessions;[m]-[n]switchesincluded...............97Table3.5.Betweenconditioncomparison;session1excluded;[m]-[n]switchesincluded98Table3.6.Differentcondition;session1excluded;[m]-[n]switchesexcluded.......................100Table3.7.Betweenconditioncomparison;allsessions;[m]-[n]switchesexcluded.............101Table3.8.Betweenconditioncomparison;session1excluded;[m]-[n]switchesexcluded.
...........................................................................................................................................................................101Table5.1.PassingratesforeachconditionandexperimentinStudy1......................................129Table5.2.Experiment1AFixedEffects.....................................................................................................134Table5.3.Experiment1Arandomeffects................................................................................................134Table5.4.Experiment1Bfixedeffects.......................................................................................................134Table5.5.Experiment1Brandomeffects.................................................................................................135Table5.6.Experiment2fixedeffects..........................................................................................................135Table5.7.Experiment2randomeffects....................................................................................................135Table5.8Numberofitemsanalyzed(wordsorsyllables),errors,errorrates,andconstraint
orderforprevioustonguetwisterexperiments..........................................................................137Table5.9.Differentcondition;allsessions;[m]-[n]swapsincluded............................................138Table5.10.NativeSharedcondition;allsessions;[m]-[n]swapsincluded..............................138Table5.11.Non-NativeSharedcondition;allsessions;[m]-[n]swapsincluded....................139Table5.12.Vowelcondition;allsessions;[m]-[n]swapsincluded...............................................139Table5.13.Differentcondition;allsessions;[m]-[n]swapsexcluded........................................139Table5.14.NativeSharedcondition;allsessions;[m]-[n]swapsexcluded..............................140Table5.15.Non-NativeSharedcondition;allsessions;[m]-[n]swapsexcluded....................140Table5.16.Vowelcondition;allsessions;[m]-[n]swapsexcluded..............................................140Table5.17.Differentcondition;session1excluded;[m]-[n]swapsexcluded.........................141Table5.18.NativeShared;session1excluded;[m]-[n]swapsexcluded...................................141
14
Table5.19.Non-NativeSharedcondition;session1excluded;[m]-[n]swapsexcluded....141Table5.20.Vowelcondition;session1excluded;[m]-[n]swapsexcluded...............................141
15
1. Introduction
Inourday-to-daylivesweencounteranenormousamountoflinguisticvariation.
Individualspeakers,forexample,widelyvaryintheirvowelproductions(e.g.,Hillenbrand,
Getty,Clark,&Wheeler,1995).Successfullynavigatingsuchwidespreadvariationrequires
ustoquicklyandeffectivelyadapt—i.e.,updatingourexpectationstobettermatchfuture
inputinagivencontextbasedonwhatwearecurrentlyexperiencinginthatcontext,soas
tobetterpredictandmoreefficientlyprocessfutureevents.1Inthecaseofphonetics,this
involvesadaptingtonovelspeakers,dialects,languages,andothertask-relevantproperties
thatservetodistinguishdifferentcontexts.Suchflexibilityiscriticaltoourabilityto
accuratelyperceivespeechfromdifferentspeakersandindifferentenvironments,aswell
ashaveourspeechbeaccuratelyperceivedbyothers.Thetypeofvariationweencounteris
notrandom,however—itishighlystructured,withindividualspeakers,dialects,languages
andcontextsallvaryingindifferentwaysandtodifferentdegrees(Kleinschmidt&Jaeger,
2015).Uncoveringtheunderlyingstructurethatgeneratesdistinctpatternsofvariationis
criticaltosuccessfuladaptation.Todoso,speakersmustusetheirpriorexperiencewith
variationasaguide,makingcausalinferencesaboutthesourceofvariation.Doingso
allowsspeakerstoadapttosystematicandrelevantvariation,whileignoringincidental
1Wedifferentiateshorter-termadaptationfromlonger-termlearningprimarilybasedonthedifferenttimecoursesforeachprocess.
16
variationnotrelevantforthetaskathand(Liu&Jaeger,2018;Samuel,Brennan,&Kraljic,
2008).
Forexample,ifsomeonehearsatalkerconsistentlyproduceanidiosyncratic[s]that
soundsunusuallylike[ʃ](e.g.,shickinsteadofsick),adaptingtothatspecificindividual’s[s]
productionswillbeadvantageousforperceivingthatindividual’sspeechinthefuture,asit
isastablepropertyoftheindividualspeaker.Thisisaformofsystematicvariation,guided
bythelistener’spastexperiencewithindividualphoneticvariation(e.g.,Kraljic&Samuel,
2007).If,ontheotherhand,thespeakerhappenstohaveapenintheirmouthwhile
talking,thelistenercaninferthatthesourceoftheidiosyncratic[s]productionmaybedue
toanincidentalfactor:theobstructionfromthepen.Thisincidentalvariationisunlikelyto
bepredictiveofthespeaker’sfuturespeechinothercontexts(i.e.,whentheydonothavea
penintheirmouth);assuch,listenersarelesslikelytoadaptundertheseconditions(Liu
andJaeger,2018;Samueletal.,2008).Critically,listenersdonotcompletelydisregardall
causallyambiguousinput(i.e.,idiosyncraticproductionswhenthetalkerhasapenintheir
mouth).Instead,theyholditinmemory,asitmaybepredictiveoffutureinputinsimilar
contexts(i.e.,futureproductionswhenthetalkerhasapenintheirmouth)oritmayprove
tobepredictiveafterfurtherdisambiguatingevidence(i.e.,thetalkerproducesthesame
idiosyncraticproductionswithoutapenintheirmouth;Liu&Jaeger,2018;Kraljic&
Samuel,2011).Inotherwords,adaptationrequireslistenerstoproperlyattributevariation
toitsunderlyingsourceforthegiventask.
Inthisdissertation,wewillfocusontheroleofsystematicvs.incidentalvariationin
adaptationtonovelphonotacticconstraints.Phonotactics—constraintsonthepossible
17
sequencesandpositionsofsoundswithinwordsandsyllables—differwidelybetween
languages,butmuchlesssobetweenindividualspeakersofasinglelanguagevariety.
English,forexample,allowsvoicedplosives(i.e.,[b],[d]and[g])insyllable-finalposition;
Dutch,ontheotherhand,onlyallowsvoicelessplosivesinsyllable-initialposition.While
suchphonotacticdifferencesbetweenspeakersofDutchandEnglisharesystematic,
encounteringtwoEnglishspeakerswhodifferinthiswayisunlikely.Thereare
communicativeconstraintsagainstwidespreadphonotacticvariationbetweenspeakers
withinlanguagevarieties,asindividualspeakersdifferinginthiswaywouldleadto
unreliablecuestowordandsyllableboundaries,resultinginfrequenterrorsinlexical
access(Pierrehumbert,2001).
Theunderlyingstructureofphonotacticvariation,andspeakers’previous
experiencewiththisvariation,likelyplaysaroleinthewaysspeakersadapttonovel
phonotacticconstraints.Researchoverthepast20yearshasfoundthatspeakersquickly
adapttonovelphonotacticconstraints(e.g.,“[s,ʃ,f]arerestrictedtoonsetposition,while
[p,t,k]arerestrictedtocodaposition”)inbothspeechproduction(e.g.speecherror
patterns;Dell,Reed,Adams,&Meyer,2000)andperception(e.g.memoryerrorpatterns;
Bernard,2015).
Inthisdissertation,weexplorethehypothesisthatphonotacticadaptationis
constrainedbythetypesofcausalinferencesspeakersmakeaboutthesourceof
phonotacticvariation.Thesecausalinferencesarebasedonspeakers’priorexperience
withphonotacticvariation:speakersofdifferentlanguagessystematicallydiffer,often
quitedrastically,intheirphonotactics;whilespeakersofthesamelanguagevarietiesare
18
unlikelytovaryinthisway.Assuch,wepredictthatwhenlearnersencountersuch
variationbetweenspeakersofasinglelanguagevariety,theywillinferitisincidental,
ratherthansystematic.Inotherwords,theywillnotattributethesourceofthevariationas
beingadurable,context-independenttraitofthetalker.Thishypothesispredictsthatwhen
speakersareexposedtomultipletalkerswithdistinctphonotacticgrammars,eitherin
perceptionorproduction,theywillshowahighdegreeofadaptationifthosetalkersclearly
differintheirlanguagebackground(e.g.,onenativeHinditalkerandonenativeEnglish
talker),andalowdegreeofadaptationiftheydonot(e.g.,twonativeEnglishtalkers).
Indeed,theonlypreviousstudytoexamineadaptationtoindividualtalkerswhosharea
languagevariety(e.g.“TalkerAdoesn’tendtheirsyllablesin/f/;TalkerBdoesn’tendtheir
syllablesin/n/”)foundthatspeakersdidnotadaptundertheseconditions,usinga
speededrepetitiontask(Onishi,Chambers,&Fisher,2002).
Thesepredictionsaretestedintwostudies.Study1examinesphonotactic
adaptationinperception,exposinglistenerstotwotalkers,eachofwhomdiffersintheir
phonotacticgrammar(e.g.,“forTalkerA,[s,ʃ,f]arerestrictedtoonsetposition;forTalker
Bwhile[p,t,k]arerestrictedtocodaposition”).Crucially,insomeconditionsthetalkers
differintheirlanguagebackgrounds;inotherconditions,thetalkerssharealanguage
background.Adaptationisassessedviaarecognitionmemoryparadigm(e.g.,Bernard,
2015;Denby,Schecter,Arn,Dimov,&Goldrick,2018).InExperiment1,listenersare
exposedtonativeFrenchandEnglishtalkersexhibitingdifferentphonotacticconstraints.
Wepredictthatwhenlistenersareexposedtotalkerswithasharedlanguagebackground
(i.e.,2Englishtalkersor2Frenchtalkers)theywillinferthetalkerssharealanguageand
19
thereforeasinglephonotacticgrammar,suggestingthephonotacticvariationisincidental,
andshowalowdegreeofadaptation.Whenlistenersareexposedtotalkerswhodifferin
theirlanguagebackground(i.e.,1Englishtalkerand1Frenchtalker),wepredictlisteners
willinferthevariationisasystematicqualityofthetalkers’languages,andtherefore
listenerswillshowahighdegreeofadaptation.
Resultsshowedthatlistenersadaptedtothenovelconstraintsinallconditions,
suggestingthattheyattendedtothedifferingphonotacticsevenwhenthesourceof
variationwascausallyambiguous(i.e.,talkerssharedalanguagebackground).Thehighest
degreeofadaptationoccurred,however,whentalkersdifferedintheirlanguage
backgrounds,andthelowestdegreeoccurredwhenbothtalkerswerenativespeakers.In
otherwords,listenersadaptedtoahigherdegreewhenthesourceofthevariationwas
causallyunambiguous(i.e.,talkersdifferedintheirphonotacticsduetothedifferencein
theirlanguagebackgrounds).Surprisingly,listenersadaptedtoamoderatedegreewhen
bothtalkerswerenon-native(i.e.,Frenchtalkers).Thismaybeduetolisteners’lackof
knowledgeofnon-nativelanguages—listenersarelikelymoreconfidentjudgingtwonative
talkersasspeakingthesamelanguagethantwonon-nativetalkers.
HowcanweunderstandlistenerbehaviorinExperiment1?Listenersmaysimplybe
sensitivetowhethertheysharethelanguagebackgroundofatalker(native)orifthetalker
doesnotsharetheirlanguagebackground(non-native).Alternatively,listenersmaybe
sensitivetotalkerlanguagebackgroundsregardlessofwhethertheythemselvessharea
backgroundwiththetalker.InthesecondexperimentofStudy1,weinvestigatethe
structureoflistenerknowledgeofnon-nativephonotacticvariation.Listenersareexposed
20
totalkersoftwonon-nativelanguages(HindiandHungarian).Iflistenersmakedistinctions
withinnon-nativephonotacticgrammars,theyshouldadaptwhentalkersdifferintheir
languagebackgrounds.If,ontheotherhandlistenersonlydistinguishbetweennativevs.
non-nativephonotactics,withoutfurtherdistinctionsbetweennon-nativephonotactics,
theywillinferbothnon-nativespeakersshareasinglephonotacticgrammarandshowa
smalldegreeofadaptation.Resultssuggestedlistenersweresensitivetodistinctions
withinnon-nativelanguages:listenersadaptedtoahighdegreewhentalkersdifferedin
theirlanguagebackgrounds,regardlessofwhetheroneofthemwasnative(e.g.,English
talkervs.Hinditalker)ornot(e.g.,Hungariantalkervs.Hinditalker).
InStudy2,weexaminewhetherspeakersmakecausalinferencesaboutphonotactic
variationinspeechproduction,usingamodifiedtonguetwisterparadigm(Dell,etal.,
2000).Participantsareexposedtomultiplemodeltalkers—nativeEnglishand/ornative
Germantalkers—exhibitingdistinctphonotacticconstraints.AsinStudy1,thesemodel
talkerseitherdifferintheirlanguagebackgroundsorshareasinglelanguagebackground;
wepredictahighdegreeofadaptationwhenmodeltalkersdifferintheirlanguage
background,andalowdegreeofadaptationwhenmodeltalkerssharealanguage
background.Recentevidencefromphonotacticadaptationinproduction,however,points
toapurelyassociativeaccountofphonotacticadaptationinproduction(e.g.,Anderson,
Holmes,Dell,&Middleton,2019),inwhichinferencesaboutthecausesofvariationarenot
alwaysintegratedintoadaptation.Assuch,itispossiblewemayfinddivergentresultsin
studies1and2duetobroaddifferencesinadaptationinspeechperceptionandproduction
(e.g.,Samuel,2011).
21
ResultsfromStudy2weredifficulttointerpretgivenasurprisinglyhighnumberof
illegalerrors,acrossallconditions,fortheconsonantsthatwerethetargetofthe
constraint.Despitethis,thereissomeevidenceofadaptationwhenmodeltalkersdiffered
intheirlanguagebackground,butnoevidenceofadaptationwhenmodeltalkersshareda
languagebackground.
Together,thesestudiesaimtoextendtheoriesoftheroleofcausalinferencein
adaptationintothedomainofphonotactics;explorepossibledifferencesinadaptation
betweenperceptionandproduction;andshedlightonthemechanismsunderlyingthe
speedandflexibilityofphonotacticadaptation.
22
2. Study1
2.1. Introduction
Listenersencounterahugeamountofvariationintheirday-to-daylinguistic
experience—forexample,men,womenandchildrenshowconsiderablevariationintheir
vowelproductions(e.g.,Hillenbrand,Getty,Clark,&Wheeler,1995)—andyetspeech
perceptionisremarkablyaccurate.Overthelast25years,theimportanceoflistener
adaptationtonoveltalkershascomeintofocus:forexample,whenlistenersareexposedto
idiosyncraticproductionsofspeechsounds,theyareabletoadapttheirphonemiccategory
boundariesaccordingly,anddosodifferentlyfortalkerswithdifferentproductions(e.g.,
Kraljic&Samuel,2007).Thisadaptationallowslistenerstonavigatetheinter-talker
variabilitytheyencounter,helpingthemtopredictspeechfromthattalkerinthefuture.
Theflexibilityofperceptualadaptation,however,isconstrainedbythetypesof
highlystructuredvariationlistenersencounter—individualtalkersdonotvaryfreely.
Beyondidiosyncraticdifferencesbetweenindividuals,talkersvaryonanumberof
linguisticandsociolinguisticdimensions(e.g.,nativevs.nonnativetalkers,dialect,age,
race;seeDrager,2010forreview).Forlistenerstousethisstructuredvariationtotheir
advantageinperception,theymustidentifythesourceofthevariation,andtheunderlying
systemthatgeneratesit(e.g.,Kleinschmidt,2018;Kleinschmidt&Jaeger,2015).Indoing
so,speakersmustdistinguishbetweensystematicvariationthatisrelevantforagiventask,
andincidentalvariationthatisirrelevantforthetask(e.g.,Kraljic&Samuel,2011;Kraljic,
Samuel,&Brennan,2008;Liu&Jaeger,2018).
23
Forexample,imagineyouarehavingaconversationwithsomeonewhohasabad
cold.Thechangestothatspeaker’svocaltractfromthecold(e.g.,occlusionofthenasal
tract)introducedistortionstotheacousticpropertiesoftheirspeech(e.g.,Tull,Rutledge,&
Larson,1996).Theacousticdistortionintroducedbythecoldistemporary,andnot
systematicforthatspeaker—itisnotapartofthespeaker’susualstate,andthusnotuseful
forperceivingthattalker’sspeechinthefuture.Ifyouencounterthisspeakeragainina
week,it’slikelythatthisacousticdistortionwillnolongerbepresent.Assuch,listeners
shouldputlittleweightonthoseproductionswhenupdatingtheirexpectationsofthat
talker’sfuturespeechwhenhealthy.Thisconstitutesoneofthecorechallengesofspeech
processing:adaptingtosystematicandrelevantvariation(e.g.,talkerdifferences)thatwill
helpyoubettercommunicateinthefuture,whiledeemphasizingirrelevantandincidental
variation(e.g.,noisyproductions)2.
Inthisstudy,weexaminethetypeofrelevantvs.irrelevantvariationspeakers
experiencewithregardstophonotactics—constraintsonpossiblepositionsandsequences
ofsoundsinwordsandsyllables—andhowthesystematicityofvariationaffects
adaptation.Englishspeakers,forexample,unconsciouslyknowthatsung[sʌŋ]isa
phonologicallylicitstructurebutngus[ŋʌs]isnot,as[ŋ]canonlyappearinsyllable-final
positioninEnglish(e.g.,Chomsky&Halle,1965).Phonotacticconstraintsdiffer2Itshouldbenotedthatwhilethephoneticvariationintroducedbythecoldisirrelevantforguidingfutureexpectationsaboutthatparticulartalker,itishighlyrelevantforsimilarcontextsencounteredinthefuture:namely,whentalkershavecolds.Assuch,wemightexpectthatlistenersbuildamentalrepresentationof“coldspeech”,reflectingthestructureofphoneticvariationtheyexperience,thatallowsthemtobettercomprehendsuchspeechinthefuture.
24
systematicallybetweenlanguages:unlikeEnglish,Vietnameseallows[ŋ]inonsetposition,
aswellascoda(e.g.,[ŋũ],“sleep”).Suchdifferencesbetweenlanguagesarecommon:
Russianallowsconsonantclusterssuchas[stv]thatarenotlegalinEnglish;Dutchdoesnot
allowvoicedobstruents,suchas[d],attheendsofwords;Hawaiiannotallowany
consonantsincodaposition;andsoon.
Twoindividualspeakersofasinglelanguagevariety,however,generallydonot
differinthisway—sharingaphonotacticgrammarispartofwhatitmeanstosharea
languagevariety.Forexample,encounteringanativeEnglishspeakerwhosegrammar
allows[ŋ]inonsetpositionisexceedinglyunlikely.Moreover,Pierrehumbert(2001)
arguesthatphonotacticconstraintsmustbewidelysharedacrosstalkerswithinalanguage
forcommunicationtobepossible:ifspeakersofasingledialectsystematicallyvariedin
theirphonotacticgrammars,phonotacticcuestowordandsyllableboundarieswouldbe
unreliable,leadingtosystematicerrorsinlexicalaccess.(Forexample,itcanbedifficultto
tellwordsapartinspeechfromanunfamiliarlanguage,dueinparttoalackofknowledge
ofthatlanguage’sphonotacticcuestowordboundaries.)
Assuch,thereisasystematicasymmetryinthedegreetowhichphonotactic
constraintsvarybetweenspeakersofdifferentlanguagesandspeakerswhosharea
languagevariety.Forexample,anEnglishspeakerandaFrenchspeakerwillvaryintheir
phonotacticgrammarsmuchmorethantwoEnglishspeakerswhosharetheInlandNorth
dialect.Weexaminehowthesesystematicdifferencesshapeadaptationtonovel
phonotacticconstraints.
25
Overthepasttwodecades,laboratoryparadigmshavebeendevelopedthatallowus
toexamineadaptationtovariationinphonotactics.Thesestudiessuggestthattalkersand
listenersadapttonovelphonotacticswithsurprisingspeed:participantsareabletolearn
arbitraryconstraints(e.g.,[n]and[f]cannotappearincodaposition)inbothspeech
production(e.g.speecherrorpatterns;Dell,Reed,Adams,&Meyer,2000)andperception
(e.g.memoryerrorpatterns;Bernard,2015;Denby,Schecter,Arn,Dimov,&Goldrick,
2018).
Wehypothesizethat,muchliketheexamplesofperceptualadaptationdiscussed
above,listenersadapttorelevantorsystematicphonotacticvariation,whileignoring
variationirrelevanttothetaskathand.Unliketalker-specificphonetics,however,
phonotacticsvarylittlebetweenindividualspeakerswithinaspeechcommunity.This
suggeststhatlistenersmaytreatdifferencesbetweenindividualtalkersasirrelevant,and
inferasingle,sharedphonotacticgrammarfortwotalkerswhoshareadialect.Indeed,one
studythatinvestigatedtalker-specificphonotacticconstraints(e.g.,forFred,stopsare
restrictedtocodaposition,andfricativesareunconstrained;viceversaforBarbara)foundno
evidenceofadaptation(Onishi,Chambers,andFisher,2002).Extendingpreviousfindings,
wepredictlittleadaptationwilloccurwhenspeakerssharealanguageevenifboth
speakersarenon-native(relativetothelistener).Inotherwords,whetherthespeakers
shareanativelanguagewiththelistenerisimmaterial—aslongasthespeakerssharea
languagewithoneanother,listenersshouldinferthetwospeakersshareaphonotactic
grammar,andthereforenotadapt.
26
Inaddition,thecomplementarypredictionhasyettobeexamined.Whenlisteners
encounterphonotacticvariationbetweentwospeakerswhodonotsharealanguage
background(e.g.,anativespeakerofEnglishvs.anativespeakerofFrench),wepredict
theyshouldtreatsuchdifferencesassystematicbasedontheirpriorexperiencewith
phonotacticvariationacrosslanguages,andinferseparatephonotacticgrammarsforeach
speaker.
Itisunclearwhattypeanddegreeofexperiencewithnon-nativelanguagesis
requiredtomakeinferencesaboutspeakerlanguagebackgroundandphonotactic
variation.It’spossiblethatlistenersonlyrequireoccasional,incidentalexposuretonon-
nativephonotactics(fromeitherspeakersofdifferentlanguages,oraccentedspeakersof
theirnativelanguage).Manylistenerswouldnaturallycomeacrosssuchspeechintheir
dailylivesinanindustrializedsocietysuchastheUnitedStates(MechanicalTurkworkers,
whichisthepopulationwesampledfrom,alsohavehigherratesofeducationthanthe
generalU.S.population;Levay,Freese,andDruckman,2016).Alternatively,listenersmay
requireahighdegreeofexposure,suchashavingspenttimelearninganon-native
language,orproficiencyintwoormorelanguages.Toaddressthisquestion,weanalyzethe
self-reportedlanguagebackgroundsofourlisteners.
Thisstudyconsistsoftwosetsofartificialgrammarexperimentsinwhichnative
Englishlistenersareexposedtotwotalkers,eachofwhomexhibitsadifferenttalker-
dependentphonotacticconstraint.Crucially,thelanguagebackgroundoftalkersis
manipulated.InExperiments1AandB,talkerssharealanguagenativetolisteners(2
Englishtalkers);shareanon-nativelanguage(2Frenchtalkers);ordonotsharealanguage
27
background(1Frenchtalker,1Englishtalker).Wepredictthatlistenerswilladaptonly
whentalkersdifferintheirlanguagebackground,aslistenersinferfromtheirprior
experiencethatlanguage-dependentvariationisrelevant,whilewithin-dialectvariationis
incidental.Topreviewtheresults,wefinddifferencesinthedegreeofadaptation,rather
thanthecategoricalpresencevs.absenceofadaptation,basedonsharedordifferenttalker
languagebackgrounds.Thisleadsustorefineourhypothesis:listenersadapttoagreater
degreetovariationtheyinterpretasrelativelysystematic,andtoalowerdegreeto
variationtheyinterpretasrelativelyincidental.
InExperiment2,weexaminethestructureoflistenerknowledgeofvariationinnon-
nativephonotactics.Listenersmaybesensitiveonlytowhetherornottheysharea
languagebackgroundwiththetalker,distinguishingonlybetweenthelistener’snative
languageandallnon-nativelanguages.Alternatively,listeners’priorknowledgecould
makedistinctionsbetweenmultiplenon-nativelanguages.Toexplorethisquestion,
listenersareexposedtonon-nativetalkerswhodifferintheirlanguagebackgrounds(1
Hinditalker,1Hungariantalker).
2.2. Background
2.2.1. Phonotactics
Knowingthephonotacticsofalanguageentailsknowingrealwordsinthatlanguage
(e.g.,Englishflick),aswellaswhatconstitutespossiblewords(frick),andwhatconstitutes
impossiblewords(fnick;Chomsky&Halle,1965).Thisknowledgeguidesperceptionin
28
profoundways,asiteliminatessomeoptionsaspossiblewordsbutnotothers.For
example,MassaroandCohen(1983)findthatthesametoken,ambiguousbetween[r]and
[l],isperceiveddifferentlybasedonthelegalityofthephonotacticcontextinwhichit’s
heard:inthe[t?i]context,it’smoreoftenperceivedas[r];inthe[s?i]context,it’smore
oftenperceivedas[l].Phonotacticsalsoinfluenceswordsegmentation(firstlanguage:
McQueen,1998;secondlanguage:WeberandCutler,2006),andanumberofother
perceptualprocesses(e.g.,Dupoux,Kakehi,Hirose,Pallier,&Mehler,1999;Otake,
Yonehama,Cutler,vanderLugt,1996;Pitt&McQueen,1998;Vitevitch&Luce,1999;fora
review,seeGoldrick,2011).
Giventheimportanceofphonotacticsinpredictionprocesses,anefficientlearner
shouldquicklyadapttonovelphonotacticconstraintstobetterguideperceptioninthe
future.Indeed,listenerscanquicklylearnartificialphonotacticconstraintsinexperimental
settings(e.g.,Bernard,2015;Denby,etal.,2018;Onishietal.,2002;Richtsmeier,2011;
Steele,Denby,Chan,&Goldrick,2015).Bernard(2015),forexample,exposedparticipants
toaseriesofspokensyllablesexhibitinganexperimentalconstraint(e.g.,[p]cannotappear
incoda;[f]cannotappearinonset).Participantswereaskedaftereachsyllablewhether
theyhadheardthatsyllableearlierintheexperiment.Afteranumberofrepetitionsofthe
exposureset,ahandfulofnovelsyllableswerepresented,halfofwhichfollowedthe
constraintandhalfofwhichviolatedtheconstraint.Participantsweremorelikelytofalse
alarmonnovelsyllablesthatfollowedtheconstraintthanthosethatviolatedit,suggesting
participantswereutilizingthenovelconstrainttomakememoryjudgments.
29
2.2.2. AdaptationandVariation
Theguidingroleofpriorexperienceisageneralpropertyofadaptation.Inthe
perceptionoffaces,forexample,learnersadaptdifferentlytonovelfaceshapesthatare
similarvs.dissimilartofacestheyhavepreviouslyexperienced(e.g.,Little&Apicella,2016;
Webster,Kaping,Mizokami,&Duhamel,2004),suggestingthatthetypeofvariation
learnershavepreviouslyexperiencedaffectshowtheyadapttonovelpatterns.Similarly,in
theadaptationtothephoneticsofindividualtalkers(e.g.,Creel,Aslin,&Tenenhaus,2008;
Eisner&McQueen,2005;Goldinger,1996;Kraljic&Samuel,2007;Nygaard&Pisoni,1998;
Pardo,2006),adaptationismotivatedbythehugeamountofinter-andintra-speaker
phoneticvariationspeakershavepreviouslyencountered(e.g.,Hillenbrandetal.,1995).
NygaardandPisoni(1998),forexample,foundthatlistenersmoreaccuratelyrecognized
wordsandsentencesinnoiseforfamiliartalkers,suggestingtheylearnidiosyncratic
featuresofthattalker’sspeechandusethatknowledgetoguideperceptionofthattalkerin
thefuture.
Inadditiontoadaptationtovariationbetweenindividualtalkers,adaptationis
conditionedonthestructuredvariationintroducedbyhigher-level(socio)linguisticfactors
(e.g.,Kleinschmidt,2018).Asubstantialbodyofworkhassuggestedthatlistenersencode
thisstructuredvariation(seeDrager,2010,forreview).Inturn,thesefactorsguide
adaptationtonovelspeakers,mostclearlyinthecaseofnon-nativeaccentadaptation
(Bradlow&Bent,2008;Reinisch&Holt,2014;Xie&Myers,2017).Forexample,fornative
Englishlisteners,exposuretoSpanish-accentedtalkersimprovesrecognitionaccuracyfor
30
novelSpanish-accentedtalkers,especiallyforwordsincludingSpanishvowelsthatareless
characteristicofEnglish(Sidaras,Alexander,&Nygaard,2009).Thissuggestslisteners
havemodelsofspecificnon-nativeaccents,encodedasdistinctfromnativeaccents,and
usethisinformationtoguideadaptation.Thereisalsoevidencethatinsomecontexts
nativelistenerstreatnon-nativespeechasadistinctconcept,astheyareabletogeneralize
acrossnon-nativeaccents(Baese-Berk,Bradlow,&Wright,2013).
Listenerscanalsousepastexperiencewithphoneticvariationtomakecausal
inferencesaboutthesourceofthevariationforthenovelexperimentaltalkersand
contexts.Kraljic,Samuel,andBrennan(2008)exposedlistenerstoatalkerproducing
ambiguous[s~ʃ]productions.Inonecondition,listenersheardtheambiguousproductions
inanexposurephasewithavideodepictingthetalkerwithapenintheirmouth;ina
secondcondition,listenerswereexposedtothesametalker,butwithavideodepictingthe
talkerholdingthepenintheirhand.Thepeninthemouthconditionprovidedlisteners
withanincidentalsourceforthevariation:thepeninthetalker’smouthwasdisrupting
theirproductions.Assuch,listenersdidnotadapttothetalker’sidiosyncratic[s~ʃ]
boundaryinthatcondition.Listenerswhosawthevideoofthetalkerwiththepenintheir
hand,however,inferredthatthevariationtheywereexposedtowasasystematic
characteristicofthattalker’sspeech,andadaptedaccordingly–appropriatelyutilizing
previousexperiencetoconstrainadaptation(seealsoKraljic&Samuel,2011;Liu&Jaeger,
2018).
Recentworkintherationallearnerframeworkhascharacterizedresultssuchas
thesebyviewingadaptationasprocessofuncoveringtheunderlyingstructurethat
31
generatesobservableeventsandinferringcausalrelationsthathelptoexplainthoseevents
(Qian,Jaeger,andAslin,2012).Suchmodelsofhowtheworldworksallowgeneralization
tonovelsituations—priorexperiencecanguideexpectationsaboutwhatwillbe
encountered,especiallyinsimilarcontexts,andhelpmakesenseofnovelexperiences.
Priorexperiencecanalsoconstrainadaptation,whenthenovelcontextisdissimilarto
thosewehaveexperiencedpreviously.Thisrigidityisanimportantfeatureofthesystem,
astotalplasticitywouldrequireinefficientlybuildinganentirelynovelmodelforevery
novelcontext.Withinthisframework,thestructuredvariabilitythatformsthebasisofour
experiencewithlanguageisencodedviaahierarchicalindexicalstructure(Kleinschmidt&
Jaeger,2015;Pajak,Fine,Kleinshmidt,&Jaeger,2016).Forexample,listenerscouldmodel
structuredphoneticvariationbyincludingdifferentlanguagesatthetopofthehierarchy
(e.g.,Hindi,French),withnativevs.non-nativeaccentsonestepbelow,followedbydialects
withinthenativeaccentandsociolinguisticgroupings(e.g.,gender),withindividual
speakersatthebottom.
2.2.3. PhonotacticAdaptationandVariation
Iflistenersbuildahierarchicalindexicalstructurebasedonthevariationthey
encounter,suchastructureshoulddifferforlevelsoflinguisticstructurethatexhibit
differentpatternsofstructuredvariation.Giventhatphonotacticvariationisgreatest
32
acrosslanguagesandsmallestacrossindividuals3,wehypothesizethatlistenerswill
assumethattheyshouldbuildseparatemodelsforspeakersofdifferentlanguages,while
theywillassumethatspeakerswithinadialectshouldbeassignedtothesamemodel.This
hypothesispredictsthatlistenersshouldshowagreaterdegreeofadaptationtotalker-
specificphonotacticswhenthetalkersdifferintheirlanguagebackground.
Thispredictionisconsistentwithevidencefrompreviousphonotacticadaptation
studies,inwhichlistenersareexposedtoartificiallanguageswithnon-nativephonotactic
constraints,andappeartoquicklylearnthesenovelconstraints,suggestingthatlisteners
assumetheyhaveencounteredanon-native“laboratorylanguage”.Indeed,participantsare
abletomaintainseparatemodelsforEnglishandalearnedlaboratorylanguage.Ina
productiontask,Warker(2013)exposedparticipantstocomplex,“second-order”
phonotacticconstraints,inwhichthepossiblepositionsofaphonemedependonfeatures
ofsurroundingphonemes(e.g.,[æ]cannotbefollowedby[s]andcannotbeprecededby[f];
thereverseconstraintistruefor[ɪ]).Participantsrequiredtwoexperimentalsessionson
separatedaystoacquirethesecond-orderconstraints(seealsoWarker&Dell,2006).
Whentheyreturnedtotheexperimentafteraweek,theyretainedtheirknowledgeofthe
experimentalconstraints,despitethehugeamountofconflictingevidenceparticipants
receivedfromEnglishintheinterveningweekbetweenexperimentalsessions(e.g.,that
3Whilephonotacticsclearlyvariesacrossdialects,theextentofthisvariationisunclear.StaumCasasanto(2008)providesevidencethatlisteners’processingofphonotacticvariantsisaffectedbyspeakerdialect,suggestingthatphonotacticscanvaryacrossdialects.Quantifyingthisvariationandexaminingitsimplicationsforphonotacticadaptationisakeyareaforfuturework.
33
[æ]canbefollowedby[s]inwordslikepass).Thissuggestslistenersmaytreatartificial
languagesasnon-nativelanguages,evenwhentheexposuretalkerhasanativelanguage
background.Asawhole,theseresultsareconsistentwiththehypothesisthatlearnerstreat
whattheylearninthelabasadistinctlanguage.
InExperiments1AandB,wetestthepredictionthatlistenerswillshowmore
robustadaptationwhentwotalkersdifferintheirlanguagebackgrounds,butnotwhen
theysharealanguagebackground.Notethatthisdistinctionispredictedonlyiflisteners
candetectthatthetwotalkersdifferinthelanguagebackgroundinthefirstplace.Assuch,
wepredictthatthedegreeofadaptationwillbeafunctionofhowmuchevidencelisteners
havethatthetwotalkersdifferinlanguagebackground.Weexaminethisbymanipulating
thestrengthofthecuetolanguagebackground.InExperiment2,weusephonotactic
adaptationtoexplorethestructureoflisteners’modelsofnon-nativephonotactics.Do
listenersmaintainmodelsofonlyanativevs.non-nativegrammar,ordotheymake
distinctionsbetweennon-nativelanguages?
2.3. Experiment1A
Inanartificiallanguageparadigm,weexposeparticipantstosecond-order
constraintsthatrequiretrackingtalkerinformation(e.g.,TalkerA’scodasarerestrictedto
[s,ʃ,f];TalkerB’scodasarerestrictedto[p,t,k]),whilemanipulatingthelanguage
backgroundofthetwotalkers.Theexperimentcontainsfourconditions:intheNative
Sharedcondition,bothtalkersarenativeEnglishspeakers;intheNon-NativeShared
34
condition,bothtalkersareFrenchspeakers;intheWeakDifferentandStrongDifferent
conditions,onetalkerisaFrenchspeaker,whiletheotherisanativeEnglishspeaker.Each
participantisexposedtoasinglepairoftalkersinabetween-participantdesign.TheWeak
DifferentandStrongDifferentconditionsaredistinguishedbythestrengthoftheacoustic
cuetotheFrenchtalker’slanguagebackground:intheStrongcondition,theFrenchtalker
producesavoweluncharacteristicofEnglish(frontrounded[y]);intheWeakcondition,
theFrenchspeakerproducesthemoreEnglish-likebackrounded[u]vowel(seeAppendix
AforacousticanalysisofFrenchvowels).Notethatboththefrontrounded[y]andback
rounded[u]Frenchvowelsareperceptuallyassimilatedto[u]bynativeEnglishlisteners
(Levy,2009);thatsaid,inboththeWeakandStrongconditions,thereareanumberofcues
totalkerlanguagebackground,astherearemanyphoneticdifferencesbetweenFrenchand
Englishbeyond[y].First,whileFrenchandEnglish[i]areacousticallysimilar(Strange,
Weber,Levy,Shafiro,Hisago,andNishi,2007),French[u]isproducedwithalowerF2(i.e.,
furtherback)thanEnglish[u],althoughthisdifferenceislikelynotaslargeasthatbetween
French[y]andEnglish[u](Flege,1987).Second,voicingdistinctionsforFrenchplosives
differfromthoseinEnglish:Frenchvoicelessplosivesareshort-lagandunaspirated(i.e.,
shortvoiceonsettime),ratherthanlong-lag(longvoiceonsettime)andaspirated,asin
English;andFrenchvoicedplosivesarefrequentlypre-voiced(negativevoiceonsettime)
ratherthanshort-lag,asinEnglish(Caramazza&Yeni-Komshian,1974).Third,coronal
consonants—particularlyplosivessuchas[t]—tendtobeproducedfurtherforwardinthe
mouth(i.e.,asdentalstops)inFrenchthaninEnglish(Dart,1998).
35
Table2.1.SummaryofconditionsinExperiment1A,alongwithexperimentalspeakerlanguagebackground,gender,andstimulusvowels.NotethatvirtuallyalllistenerswerenativeEnglishspeakers.
NativeShared Non-NativeShared
WeakDifferent
StrongDifferent
SpeakerLanguageBackground
SharedEnglish
SharedFrench
Englishvs.French
Englishvs.French
Englishvs.FrenchVowels
[i,u][i,u]
[i,y][i,y]
[i,u][i,u]
[i,u][i,{u/y}*]
SpeakerGender Different Different Same SamePredictedDegreeofAdaptation
Low Low Moderate High
*N.B.ForStrongDifferentcondition,Frenchsyllablesincluded[y]infamiliarizationsyllablesand[u]ingeneralizationsyllables.
Iflistenersadaptbasedontheirpriorexperiencewithphonotacticvariationand
talkerlanguagebackground,theyshouldadapttoagreaterdegreeintheDifferent
conditions,sincetalkerswhodifferinlanguagebackgroundaremorelikelytodifferin
theirphonotacticgrammars.AmongtheDifferentconditions,thetwotalkersare
phoneticallylessdistinctintheWeakcondition;assuch,listenershavelessevidencethat
thetwospeakersdonotsharealanguagebackground.Thuswepredictagreaterdegreeof
adaptationfortheStrongDifferentconditionthantheWeakDifferentcondition.Inboth
Sharedconditions,talkersdonotdifferinlanguagebackground;listeners’priorexperience
shouldsuggestthatthetalkerswillunlikelydifferintheirphonotacticgrammars.Assuch,
wepredictthesmallestdegreeofadaptationintheseconditions.
Participantsaretestedusingacontinuousrecognitionmemorytask(Bernard,2015,
2017;Denby,etal.,2017;Steele,etal.,2015),inwhichtheyareauditorilypresentedwitha
seriesofsyllablesandaskedwhethertheyhavepreviouslyheardeachsyllablewithinthe
36
experiment.Participantsarefirstexposedtomultiplerepetitionsofasetoffamiliarization
syllables,allofwhichfollowthephonotacticconstraint(e.g.,SpeakerAsaysfut;SpeakerB
sayspuf).Afterthefirst4repetitionstothefamiliarizationsyllables,listenershear9more
repetitionsoftheentiresetoffamiliarizationsyllables,butnowwithahandfulofnovel
generalizationsyllablesmixedin.Halfofthesearelegal(i.e.,followthephonotactic
constraint),whiletheotherhalfareillegal(i.e.violatetheconstraint;forexample,Speaker
Asayingtish;SpeakerBsayingtuk).Iflistenersaretrackingtheconstraint,generalization
syllablesthatfollowtheconstraintshouldseemmorefamiliarthanthosethatdonot;as
such,participantsshouldbemorelikelytoincorrectlybelievetheyhadpreviouslyheard
legalgeneralizationsyllables.Forexample,aparticipantmighthearSpeakerAsayfut,kit,
sik,tup,etc.,multipletimesduringfamiliarization.Ifthatparticipantistrackingthe
constraint,duringgeneralizationtheymaybelievetheyhadpreviouslyheardtut,since
syllableswithsimilarphonotacticpatterns(i.e.,voicelessstopsincodaposition)appeared
infamiliarization.Incontrast,participantsshouldbeunlikelytofalsealarm(i.e.,incorrectly
respond“yes”)totus,however,sincenosyllablesspokenbyTalkerAinfamiliarization
containedcodafricatives.
Notethatspeakergenderwasalsomanipulatedacrossconditions:intheShared
conditions,speakersdifferedingender,whileintheDifferentcondition,speakersshareda
gender.Muchlikeaccent,genderconveyssociolinguisticdifferencesbetweenspeakers
(e.g.,Oh,2011).Thisservedasacontrolonphoneticandsocialdistancebetweentalkersin
eachcondition:whiletalkersintheDifferentconditionsweredistinguishedbytheiraccent,
talkersintheSharedconditionsweredistinguishedbytheirgender.Assuchineach
37
conditionthetwotalkersdifferedalongsocialandphoneticlines,eitherbygenderor
accent.
Aninitialpilotstudy(seeAppendixB)wasruntoapproximatethenumberof
participantsneededforrequisitestatisticalpower(seeAppendixCfordetailsofpower
analysis).Thedesignandanalysisoftheexperiment—includingpredictions,numberof
participants,stimulusdesign,andmodelstructure—weredefinedbeforedatacollectionin
apre-registrationontheOpenScienceFoundationplatform(osf.io/dbcqx/).
2.3.1. Participants
Basedonapoweranalysis(seeAppendixCfordetails),256participants,split
evenlybetweenthe4conditions(64percondition),wererequired.However,participants
hadtopassasetofexperimentalcriteria(seeDataAnalysissection)toensurethatthey
wereadequatelyattendingtothetask.Assuch,participantswereiterativelyrecruiteduntil
therewere64participantswhopassedthecriteriaineachcondition.Atotalof455
participantswererecruitedthroughAmazonMechanicalTurk(AMT;Buhrmester,Kwang,
andGosling,2011);ofthese,260(57.1%)passedthecriteria.Thispassingratewassimilar
topreviousstudiesusingthisparadigmoverAMT(seeSteeleetal.,2015andDenbyetal.,
2018;seeAppendixDforfullbreakdownofparticipantpassingrates).Duetolimitations
withinouronlineframeworkandAMT,4participantswhopassedthecriteriawere
exposedtoauniqueexperimentallistthatapreviousparticipanthadbeenexposedto.
Threeoftheseparticipantswereexcluded;onesuchparticipantwasincluded,however,in
38
theWeakDifferentcondition,asoneuniqueexperimentallistdidnothaveaparticipantdue
toexperimentererror.ParticipantswererequiredtohaveU.S.IPaddresses,andwere
fluentspeakersofEnglish;98.4%ofparticipantswhopassedthecriteriaself-identifiedas
nativespeakersofEnglish,while2participantsself-identifiedasspeakinganon-North
AmericandialectofEnglish.99.2%ofparticipantswhopassedthecriteriahadnospeechor
hearingimpairments.(Notethatmodelresultswerequalitativelyidenticalwhennon-
nativeandhearing-andspeech-impairedparticipantswereexcludedfromtheanalysis.)
2.3.2. Stimuli
Stimuliwererecordedinasoundproofboothata44.1kHzsamplingrate,and
normalizedto60dBSPL.4talkersrecordedstimuli:afemalenativeEnglishspeaker;a
malenativeEnglishspeaker;afemalenativeFrenchspeaker;andamalenativeFrench
speaker.Talkersproducedsyllablesfromorthographicrepresentationsofsyllablesona
monitor;orthographyreflectedthelanguagebackgroundofthespeaker.BothFrench
talkersweremultilingual,butwereinstructedtoproducethesyllablesasthoughtheywere
French,ratherthanEnglish,words.Syllableswerepresentedinarandomorder.
Stimuliconsistedofconsonant-vowel-consonantsyllableswithvoicelessstops[p,t,
k]andvoicelessfricatives[f,s,ʃ]asonsets.VowelsfortheEnglishspeakerareeither[i]or
[u].FortheFrenchspeakers,vowelsare[i]and[u]intheWeakDifferentcondition;inthe
StrongDifferentcondition,vowelsinfamiliarizationsyllableswere[i]and[y];for
39
generalizationsyllablesvowelswere[i]and[u].Theresultwasatotalof108possible
syllables(6onsetconsonants*3vowels*6codaconsonants)recordedbyFrench
speakers,and72possiblesyllables(6onsets*2vowels*6codas)recordedbyEnglish
speakers(asEnglishspeakersonlyproduced[u]andnot[y]).Participantswereexposedto
72uniquesyllablesineachcondition.
2.3.3. Procedure
Participantswereaskedtofilloutademographicformthatincludedinformation
abouttheirlanguagebackground,geographicareasinwhichtheyhadpreviouslyresided,
whethertheywereanativeornon-nativespeakerofEnglish,andwhethertheyhadany
hearingorlanguageimpairments.Participantswerefreetooptoutofansweringany
questions.
Toensurelistenershadaworkingaudioset-upandbasicfluencywithEnglish,an
audiopre-testwasadministeredinwhichlistenersidentified2Englishwordsspokenbya
talkernotinvolvedintherestoftheexperimentbytypingthewordswiththeirkeyboards.
Participantsperformedarecognitionmemorytask.Thequestion“Haveyouheard
thisbefore?”wasonthescreenfortheentireexperiment.Oneachtrial,anauditory
stimuluswaspresented.Participantsansweredthequestionbyclickinga“Yes”or“No”
buttononthescreen.Aftereachclick,therewasa500msinterstimulusintervalbeforethe
followingstimulusplayed.The“Yes”and“No”buttonsdisappearedfromthescreenuntil
thestimuluscompletedplaying.Participantshadunlimitedtimetoanswerthequestion,
40
andnofeedbackwasprovided.Therewerenobreaksinbetweenexperimentalblocks;
blockswerenotdemarcatedinanywaytotheparticipant.
2.3.4. Design
Stimuliweresplitinhalf,intogeneralizationandfamiliarizationsyllables(36each),
byonset-vowelpairs,andcounter-balancedacrossparticipants.Forexample,ParticipantA
hearstheonset-vowelpair[tu_]infamiliarizationsyllables(e.g.,toof)andtheonset-vowel
pair[ti_]innovelgeneralizationsyllables(e.g.,teef);theconversepatternholdsfor
ParticipantB(e.g.,[ti_]infamiliarization;[tu_]ingeneralization).Onset-vowelpairswere
[ti],[hi],[su],[pi],[ku],[fu]foronepattern,and[tu],[hu],[si],[pu],[ki],[fi]fortheother.
Amongthe36familiarizationsyllables,half(18)willendinfricatives,whilehalfwillendin
stops.Thesesubsetsof18syllableswilleachberepeatedbyadifferenttalker,suchthata
giventalkerwillonlyrepeatsyllablesendingineitherfricativesorstops.Thus,during
familiarizationparticipantswillbeexposedtoaphonotacticconstraintlinkingmannerin
codaposition(fricativevs.stopcoda)andspeaker.Whichtalkerproduceswhichsetis
counterbalancedacrossparticipants.Amongthe36generalizationsyllables,eachspeaker
produceshalf(18)oftheset.Amongthissubset,half(9)followtheconstraintestablished
inthefamiliarizationset,andhalfviolatethisconstraint(i.e.,bothspeakerssaynovel
generalizationsyllablesthatendinbothstopsandfricatives).
Thefirst4blocksoftheexperimentconsistsofthefamiliarizationphase.Ineach
block,participantsareexposedtothe36familiarizationsyllables(halfsaidbyeach
41
speaker)inrandomorder.Inthegeneralizationphase,thereare9furtherrandomly
orderedrepetitionsofthefamiliarizationset,buteachrepetitionisnowintermixedwith4
generalizationsyllables.Thisresultsinatotalof504trials(36familiarizationsyllables*13
blocks+36generalizationsyllables).
InbothoftheSharedconditions,thetwotalkershavedifferentgenders(e.g.,Male
EnglishandFemaleEnglishtalkerinNativeShared).IntheDifferentconditions,however,
talkershavethesamegender;talkergenderwascounter-balancedacrossparticipant(e.g.,
ParticipantAhearsafemaleFrenchtalkerandafemaleEnglishtalker;ParticipantBhears
amaleFrenchtalkerandamaleEnglishtalker).
2.3.5. DataAnalysis
Followingpreviouswork(Denbyetal.,2017;Steeleetal.,2015),participantshadto
passasetofcriteriatoensurethattheywereadequatelyattendingtothetask:asin
previousstudies,duringthegeneralizationphase(blocks5-13),participantsmustcorrectly
acceptatleast90%ofthesyllablestheyhadpreviouslyheard,andcorrectlyrejectatleast
10%ofthenovelgeneralizationsyllablesthattheyhadnotheard(notethatlooseningthe
criteriatoincludeagreaternumberofparticipantsdoesnotqualitativelyaltertheresults;
seeAppendixD).Participantswhodidnotpassthesecriteriawereexcludedfromthe
42
analysis.Inaddition,forparticipantswhoattemptedtocompletetheexperimentmultiple
timeswithinasinglesession,onlythefirstattemptwasincluded.4
Generalizationdatawasanalyzedusinglogisticmixed-effectsregressionswith
maximaleffectsstructures(Barr,Levy,Scheepers,andTily,2013).Thedependentmeasure
wastherateatwhichparticipantsfalsealarmed(e.g.,incorrectlyresponded“yes”tonovel
syllables).Fixedeffectsforthemodelconsistedoflegalityandthreecontrast-codedterms:
languagedifference,inwhichtheSharedandDifferentconditionswerecontrasted;
strength,inwhichtheWeakandStrongDifferentconditionsarecontrasted;andaccent,in
whichthetwoSharedconditionsarecontrasted.Inaddition,aninteractiontermwas
includedbetweenlegalityandeachofthecontrast-codedterms.Randomeffectsincluded
randominterceptsandrandomslopesbylegalityforbothparticipantsanditems.5Finally,a
likelihoodratiotest,betweenmodelswithandwithouteachcontrasttermasafixedeffect,
wasincludedtotestforstatisticalsignificance.
Wemeasurethedegreeofadaptationusingthesizeofthelegalityadvantage:the
“yes”responseratestolegalgeneralizationsyllablesminusthe“yes”responserateto
illegalsyllables.Ouraccountpredictsthatlistenersadaptwhentheirpriorexperience
suggeststhatthetwotalkersarelikelytohavedifferentphonotacticgrammars.Thisshould4Therewereanumberofreasonswhyparticipantswouldattempttheexperimentmultipletimes(e.g.,iftheyaccidentallyclosedtheirbrowserwindowhalfwaythrough).5Notethatrandominterceptsandrandomslopesforparticipantsanditemsbylegalitywasthemaximaleffectsstructure.Itemsweredefinedasindividualtokensspokenbyspecifictalkers(e.g.,Frenchmaletalker’s[tif]),ratherthanmoreabstract“phonological”syllables(e.g.,/tif/spokenbyalltalkers).Assuch,participantsindifferentconditionswereexposedtoadifferentsubsetofsyllables,andcontrasttermscouldnotbeincludedasrandomslopesforitems.
43
yieldaninteractionbetweenlegalityandthelanguagedifferenceterms,suchthatthe
legalityadvantageislargerintheDifferentconditions(i.e.,whentalkersdifferintheir
languagebackgrounds)thanintheSharedconditions.Further,asadaptationrequiresthat
listenersrecognizethetalkersashavingdifferentlanguagebackgrounds,wepredictthe
legalityadvantagewillbelargerwhenthecuetolanguagebackgroundisstronger(i.e.,
morerobustadaptationintheStrongDifferentconditionthantheWeakDifferent
condition),asshownbyaninteractionbetweenlegalityandcue.Finally,listenerbehavior
shouldnotchangebetweenthetwoSharedconditionsdependingonwhetherthetalkers
arenativeornon-nativespeakers.InbothSharedconditions(i.e.,twoFrenchtalkersortwo
Englishtalkers)thetalkerssharealanguagebackground,andlistenersshouldtherefore
infertheyshareaphonotacticgrammar.Assuch,wepredictnointeractionbetweenthe
legalityandaccentcontrastterm.
2.3.6. Results
A95%confidenceinterval(CI)foreachanalysisofmeanvalueswasestimatedusing
abootstrapmethod,inwhichthedistributionofastatisticisestimatedbyrepeatedly
resamplingfromtheobserveddata(withreplacement).Distributionsformeansacross
participantswereestimatedwith1,000replicates,samplingacrossmeanswithineach
participant.
Participantscorrectlyacceptedameanof91.0%offamiliarizationsyllables(CI
[90.6%,91.3%]);participantsfalselyrecognized(i.e.,incorrectlyresponded“yes”to)
44
55.9%ofnovelgeneralizationsyllables(CI[53.2%,58.6%]).Thecrucialmeasure,however,
wasthedifferenceintherateoffalserecognitionsforlegalvs.illegalsyllables,andwhether
this“legalityadvantage”wasmodulatedbytalkerlanguagebackground.Themeanlegality
advantageacrossparticipantswas12.5%(CI[10.5%,14.7%]),replicatingpreviousresults
showingthatlistenersshowhigherfalserecognitionratesonnovellegalsyllables(i.e.,
syllablesfollowingconstraintsthey’vebeenpreviouslyexposedto)thannovelillegal
syllables.Moreover,thelegalityadvantageismodulatedbylanguagebackground—ascan
beseeninFigure2.1,thelegalityadvantageismodestintheNativeSharedcondition,and
relativelylargeintheotherthreeconditions.
Figure2.1.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment1A.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment1A.Inbothpanels,errorbarsreflectbootstrapped95%confidenceintervals.
45
Ouranalysisshowedasignificantmaineffectoflegality(β=0.64,s.e.β=0.06,χ2(1)
=73.6,p<.0001),aslistenersweremorelikelytofalselyrecognizelegalsyllablesover
illegalsyllables.Inaddition,therewasasignificantinteractionoflegalitywiththelanguage
differencecontrastterm(β=0.72,s.e.β=0.20,χ2(1)=12.6,p<.001),aslistenersshowed
agreaterlegalityadvantageintheDifferentconditionsthanintheSharedconditions.
Legalityalsointeractedwithaccent(β=0.45,s.e.β=0.14,χ2(1)=9.8,p<0.01),butnot
strength(β=-0.05,s.e.β=0.14,χ2(1)=0.14,p=.70).Inotherwords,thelegality
advantagewasgreaterintheNon-NativeSharedconditionthantheNativeShared
condition,butwasnotdifferentacrosstheStrongandWeakDifferentconditions.(Seefull
modelresultsinAppendixE.)
2.3.7. Experiment1ADiscussion
Experiment1Aexposedlistenerstotalker-specificphonotacticconstraintswhile
modulatingthelanguagebackgroundoftalkers.Listenerswereabletosuccessfullyadapt
withineachcondition,acquiringtalker-specificconstraints.Moreover,thisadaptationwas
modulatedbythelanguagebackgroundofthetalkers:listenersshowedmodestdegreeof
adaptationwhenexposedtotwotalkerswithasharednativelanguagebackground(Native
Sharedcondition),andagreaterdegreeofadaptationifeitherorbothtalkershadanon-
nativelanguagebackground(StrongDifferent,WeakDifferent,andNon-NativeShared).
Therewasnodifferenceinadaptationbasedonthestrengthofthecuetolanguage
46
background(StrongDifferentvs.WeakDifferent),suggestingthatevenwiththeweakercue
tothenon-nativelanguagebackgroundoftalkers(i.e.,theFrench[u]vowel,ratherthan
[y]),listenersareconfidentofthenon-nativelanguagebackgroundofthetalker.
Countertoourpredictions,however,adaptationwasaffectedbylanguage
backgroundevenwhenbothtalkerssharedalanguage:therewasagreaterdegreeof
adaptationwhenbothtalkerssharedanon-nativelanguagebackgroundthanwhenthey
sharedanativelanguagebackground.Perhapsmoresurprising,adaptationwasequally
robustwhentalkerssharedanon-nativelanguagebackgroundaswhentheirlanguage
backgroundsdiffered(i.e.,onenativeandonenon-nativetalker).Itispossiblethatany
inclusionoftalkerswithanon-nativelanguagebackgroundincreaseslistenerconfidence
thattalkersarespeakingtwodifferentlanguages.Thismaybebecauseoftheasymmetryin
listenerknowledgeoftheirnativephoneticsvs.non-nativephonetics:duetolisteners’
extensiveknowledgeoftheirnativelanguage,whentheyencountertwonativespeakers
theyarelikelyconfidentthatthosetwospeakerssharealanguage(evenwhentheyare
bothspeakinganartificial,non-nativelanguage).Whenlistenersencountertwotalkers
withasharednon-nativelanguagebackground,ontheotherhand,theymaybeless
confidentthatthesetalkerssharealanguagebackground,giventheirrelativepaucityof
experiencewithnon-native(inthiscase,French)phonetics.
Iftheasymmetryinlistenerknowledgebetweennativeandnon-nativephoneticsis
drivingthedifferencebetweenthetwoSharedconditions,however,thisasymmetryshould
alsoresultinthegreatestdegreeofadaptationfortheDifferentconditions,whichwasnot
47
thecase.Thatis,listenerconfidenceofhavingencounteredmultiplelanguagesshouldbe
highestwhenoneofthoselanguagesisanativelanguage.
ThereweretwolimitationsofExperiment1Athatmayhaveaffectedadaptation.
First,theproductionsofthetwoFrenchtalkersshowedmarkedlydifferentpitchcontours,
tothepointthatlistenersmayhaveinferredthatthetwotalkersdidnotsharealanguage
background,increasingthelegalityadvantageintheNon-NativeSharedcondition.Thismay
havebeenduetodifferencesduringrecording,orthedifferentbackgroundsofthetwo
talkers.ThemaleFrenchspeakerwas23yearsold,andhadlivedintheUnitedStatesfor
lessthanayear.HewasfromParis,andself-identifiedasspeakingastandarddialectof
French.ThefemaleFrenchspeakerwas41yearsold,hadlivedintheUnitedStatesfor13
years,wasfromsouthofFrance,andidentifiedasspeakinganon-standarddialectof
French.
Toaddressthislimitation,inafollow-upexperimentreplicating3ofthe4
conditionsinExperiment1A(seebelow),werecordedanovelfemaleFrenchspeaker,
whoselanguagebackgroundwasmoresimilartothatofthemaleFrenchspeaker,andwho
wasinstructedtoimitatethemalespeaker’sproductionstoensurephoneticsimilarity
acrossspeakers.Assuch,wepredictalowerdegreeofadaptationintheNon-NativeShared
conditioninExperiment1Bthanin1A.
AsecondlimitationofExperiment1AwasthatintheStrongDifferentcondition,
listenerswereexposedtofamiliarizationsyllablesthatincludedtheuncharacteristic
French[y]vowel;generalizationsyllables,however,hadtheFrench[u]vowel.Thismeant
thegeneralizationsetswereidenticalacrossStrongandWeakconditions,allowingfora
48
moredirectcomparison.However,itmayhavealsoattenuatedadaptationintheStrong
Differentcondition,giventhat[u]isaweakercuetotalkerlanguagebackground.Moreover,
itincreasedthephoneticdistancebetweenfamiliarizationandgeneralizationsets,as
listenersencounteredanovelFrenchvowelinthegeneralizationsetthatwasnotpresent
infamiliarizationsyllables.LowfalserecognitionratesforsyllablesintheStrongDifferent
conditionspokenbyaFrenchtalkerandcontaining[u](38.1%)reflectedthis.Thisislower
thanfalserecognitionforFrenchsyllablescontaining[i](63.7%)intheStrongDifferent
condition,aswellasFrenchsyllablescontaining[u]intheWeakDifferentcondition
(60.0%).
Toaddressthislimitation,intheStrongDifferentconditionofExperiment1B,
familiarizationandgeneralizationsyllablesspokenbyFrenchtalkerscontainedmatching
vowels.IftheincreasedphoneticdistanceintheStrongDifferentconditiondepressedthe
legalityeffectforthatcondition,wewouldpredictagreaterdegreeofadaptationforthe
StrongDifferentconditioninExperiment1Bthanin1A.
2.4. Experiment1B
InExperiment1A,participantsadaptedtoanunexpectedlyhighdegreeintheNon-
NativeSharedconditionrelativetotheDifferentconditions.InExperiment1B,threeofthe
conditionsfromExperiment1Awerereplicated(StrongDifferent,WeakDifferent,andNon-
NativeShared)whiletwolimitationsofthepreviousexperimentwereaddressedthatmay
havecausetheunexpectedresults.
49
2.4.1. Participants
AsinExperiment1A,participantswereiterativelyrecruitedfromAMTuntilthere
were64participantsineachofthethreeconditionswhopassedtheexperimentalinclusion
criteria.Atotalof418participantswererecruited,ofwhich192(46.4%)passedthe
criteria.Noparticipantswereexcludedduetoexposuretopreviouslyseenexperimental
lists.98.9%ofparticipantswhopassedthecriteriaidentifiedasnativeEnglishspeakers.All
participantsidentifiedashavingnospeechorhearingimpairments.Noparticipant
identifiedasspeakinganon-AmericandialectofEnglish.(Notethatmodelresultswere
qualitativelysimilarwhennon-nativeparticipantswereexcludedfromtheanalysis.)
2.4.2. Stimuli
Stimulifrom3ofthe4talkerswereidenticaltothatinExperiment1A;however,
stimulifromanovelfemaleFrenchspeakerwererecordedtoreplacethestimuliofthe
femaleFrenchspeakerfromExperiment1A.Inasoundproofbooth,thenovelfemale
FrenchspeakerheardeachofthemaleFrenchspeaker’sproductionsinrandomorderover
headphones.Afterthemalespeaker’sproductionwasplayed,shewasinstructedtoimitate
it;eachsyllablewasalsoprovidedinFrenchorthography,andappearedonamonitorafter
theaudiohadfinishedplaying.
ThenovelfemaleFrenchspeakerwas24yearsold,grewupinthesouthwest
ofFrance,livedinParisasanadult,andhadlivedintheUnitedStatesforlessthanayearat
50
thetimeofrecording.Sheself-identifiedasspeakingastandarddialectofFrenchasan
adult,despitehavinggrownupspeakinganon-standarddialect(SouthwesternFrench).
IntheStrongDifferentconditioninExperiment1B,vowelsspokenbyFrench
speakersinbothfamiliarizationandgeneralizationsyllableswerealways[i]or[y].(This
differedfromtheStrongDifferentconditioninExperiment1A,inwhichFrenchspeakers
used[i]and[y]infamiliarizationsyllables,but[i]and[u]ingeneralizationsyllables.)
StimuliwereotherwiseidenticaltothoseinExperiment1A.
2.4.3. DataAnalysis
Significancewasassessedusingalogisticmixed-effectsregressionidenticaltothat
inExperiment1A,withtheexceptionofafixedeffectforaccent,whichwasnotincluded
(therewasnoNativeSharedconditioninExperiment1B).Themodelhadfixedeffectsof
legality,languagedifference(i.e.,Non-NativeSharedvs.bothDifferentconditions)and
strength(i.e.,Weakvs.StrongDifferentconditions).Aninteractiontermwasincluded
betweenlegalityandbothcontrast-codedterms;randomeffectsincludedrandom
interceptsandrandomslopesbylegalityforbothparticipantsanditems.
WepredictasignificantdifferencebetweentheDifferentconditionsandtheNon-
NativeSharedcondition,asshownbyaninteractionbetweenlegalityandthelanguage
differenceterms.Wealsopredictasignificantinteractionbetweenlegalityandthestrength
contrastterm.
51
2.4.4. Results
Participantscorrectlyacceptedameanof90.3%offamiliarizationsyllables(CI
[89.7%,90.9%]);participantsfalselyrecognized(i.e.,incorrectlyresponded“yes”to)
60.2%ofnovelgeneralizationsyllables(CI[57.3%,63.1%]).Themeanlegalityadvantage
acrossparticipantswas19.4%(CI[17.1%,21.8%]).Critically,thelegalityadvantageis
modulatedbylanguagebackground—ascanbeseeninFigure2.2,thelegalityadvantageis
moderateintheNon-NativeSharedcondition,andlargeintheDifferentconditions(the
NativeSharedconditionfromExperiment1Awasincludedforreference).
Figure2.2.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment1B.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment1B.Inbothpanels,errorbarsreflectbootstrapped95%confidenceintervals.
Alogisticmixedeffectsregressionfoundasignificantmaineffectoflegality(β=
1.02,s.e.β=0.07,χ2(1)=127.6,p<.0001),aswellaslanguagedifference(β=-1.06,s.e.β=
52
0.19,χ2(1)=28.82,p<.0001),aslistenersweremorelikelytofalselyrecognizelegal
syllables,aswellassyllablesintheDifferentconditions.Inaddition,therewasasignificant
interactionoflegalitywiththelanguagedifferencecontrastterm(β=0.57,s.e.β=0.18,χ2
(1)=9.81,p<.01),aslistenersshowedagreaterlegalityadvantageintheDifferent
conditionsthanintheSharedcondition.Legalitydidnotinteractwithstrength(β=-0.02,
s.e.β=0.15,χ2(1)=0.03,p=.87),asthelegalityadvantagewasnotsignificantlydifferent
acrosstheStrongandWeakDifferentconditions.
2.4.5. Discussion
Experiment1Breplicatedtheadaptationtotalker-specificphonotacticconstraints
foundinExperiment1A,withlistenersadaptingineachcondition.Moreover,listeners
adaptedtoagreaterdegreewhentalkersdifferedintheirlanguagebackground(Different
conditions)thanwhentheysharedanon-nativelanguagebackground(SharedNon-Native
condition),unlikeinExperiment1A.Thisprovidesevidencethatthedifferenceinlanguage
backgroundbetweentalkersiscritical,asopposedtothesimplepresenceofnon-native
talkers.
WefurtherpredictedthatthechangesinstimulusdesigntoExperiment1Bwould
resultin(a)anincreaseinthelegalityadvantageforStrongDifferentconditiondueto
consistentvowelsacrossgeneralizationandfamiliarizationsyllables,and(b)adeclinein
thelegalityeffectfortheNon-NativeSharedconditionduetotheincreasedphonetic
similarityofthetwoFrenchtalkers.WhilethelegalityadvantagefortheStrongDifferent
53
conditiondidincreaseacrossExperiments1Aand1B(fromameanof15.7%to23.1%),a
similarincreasewasfoundintheWeakDifferentcondition(from16.1%to22.7%),
suggestingthechangeinthedesignoftheStrongDifferentconditionwasnotthecauseof
thisincrease.Inaddition,fortheNon-NativeSharedcondition,thelegalityeffectwas
roughlyequivalentacrossExperiments1Aand1B(ameanof13.6%in1Aand12.7%in
1B),countertoourprediction.
TheinclusionofanovelfemaleFrenchspeakerinExperiment1Bappearstoaccount
fortheincreasedlegalityadvantageintheDifferentconditions:listenerswhoheardtwo
malespeakersintheDifferentconditionsshowedasimilarlegalityadvantageacrossthe
twoexperiments(ameanof20.2%in1Aand19.3%in1B);listenerswhoheardtwofemale
speakers,however,showedasubstantiallyhigherlegalityadvantageinExperiment1B(a
meanof25.8%)thanin1A(11.6%;seeAppendixDfortalkermeansineachconditionand
experiment).
WhydidtheinclusionofthenovelfemaleFrenchspeakerincreaseadaptationin
bothDifferentconditions,withoutloweringadaptationintheNon-NativeSharedcondition?
OnepossibilityisthattheoriginalfemaleFrenchspeakerinExperiment1Awasnot
sufficientlyphoneticallydistinctfromthefemaleEnglishspeaker,perhapsduetothe
Frenchspeaker’sextendedtimeintheUnitedStates.Ifthiswerethecase,listenerswould
haveinferredthattheysharedalanguagebackground,resultingindecreasedratesof
adaptation.Notethatthisdifferedfromouroriginalprediction:thatthefemaleFrench
speakerinExperiment1AwasnotsimilarenoughtothemaleFrenchspeaker,resultingin
higherthanexpectedadaptationintheNon-NativeSharedcondition(wesawsimilar
54
adaptationforthisconditioninExperiments1B).Asecondpossibilityisthatthefemale
Frenchspeaker’sanomalouspitchcontoursweredistracting,shiftinglistenerattention
awayfromthesegmentalleveldifferencesbetweenspeakers,andtowardstheprosodic
differences.
TheincreaseofadaptationintheDifferentconditions,whateverthecause,suggests
thatthelow-levelphoneticpropertiesoftalkers,andthedifferencesorsimilarities
betweentalkers,affectlistenerinferencesabouttalkerlanguagebackground.Replicating
thisexperimentwithnoveltalkerpairsandlanguagesisnecessarytoensurethatthe
patternofadaptationfoundinExperiment1was,infact,spurredbydifferencesinlanguage
background,ratherthantheresultofarbitraryindividualvariation.
Finally,asinExperiment1A,therewasnodifferenceinadaptationbasedonthe
strengthofthecuetolanguagebackground,providingfurtherevidencethatthe“weak”cue
issufficientforlistenerstodetectthetalker’slanguagebackground.
2.4.6. Experiment1conclusion
Experiments1AandBprovideevidencethatlisteners’previousexperiencewith
phonotacticvariation—thatspeakerswithdifferentlanguagebackgroundsexhibitalarge
degreeofphonotacticvariation,whilespeakerswithinaspeechcommunitydonot—
constrainsadaptationtonovelphonotacticconstraints.Further,theresultssuggestthat
listenershaveastructuredmodelofphonotacticvariation,withnativeandnon-native
languageseachhavingseparatephonotacticgrammars.Thespecificityoflistener
55
knowledgeofnon-nativephonotactics,however,isunclear.InExperiment2,weinvestigate
whetherlistenersassigndifferentphonotacticgrammarstodifferentnon-nativelanguages
(aswellastheirnativelanguage).Alternatively,listenersmayonlyassignasingle
phonotacticgrammartotheirnativelanguage,6andasinglephonotacticgrammartoall
non-nativelanguages.
2.5. Experiment2
InExperiment2,usingasimilardesignandrecognitionmemoryparadigmtothatin
Experiments1AandB,weexposelistenerstotwotalkers,eachofwhomexhibitsa
differentnovelphonotacticpattern.Weconceptuallyreplicate2conditionsofExperiments
1AandBusingnovelstimuli,speakers,andlanguages(HindiandHungarian).IntheMixed
DifferentconditionlistenersareexposedtoonenativeEnglishspeaker,andonenon-native
speaker(eitherHindiorHungarian),broadlyreplicatingthedesignoftheDifferent
conditionsinExperiment1.IntheNon-nativeSharedcondition,listenersareexposedto
twonon-nativespeakerswhosharealanguagebackground(either2Hindispeakersor2
Hungarianspeakers).Toaddressthestructureoflistenerknowledge,weincludeanovel
condition:intheNon-nativeDifferentcondition,listenersareexposedtotwonon-native
speakerswhodifferintheirlanguagebackground(oneHindispeakerandoneHungarian
speaker).
6Inthecaseofmultilingualspeakers,onephonotacticgrammartoeachoftheirnativelanguages.
56
Table2.2.SummaryofconditionsinExperiment2,alongwithexperimentalspeakerlanguagebackground,gender,andstimulusvowels.NNstandsfornon-native.NotethatvirtuallyalllistenerswerenativeEnglishspeakers.
Non-NativeShared
MixedDifferent
Non-NativeDifferent
LanguageBackground
SharedHindiorHungarian
Englishvs.(HindiorHungarian)
Hindivs.Hungarian
Gender Different Different DifferentPredictedDegreeofAdaptation
Moderate High WithinNN:HighNativevs.NN:Moderate
Thewithinnon-nativedistinctionshypothesisandthenativevs.non-native
hypothesismakeidenticalpredictionsintheNon-nativeSharedcondition—moderate
adaptation,followingtheresultsofExperiment1—andtheMixedDifferentcondition—a
highdegreeofadaptation.Replicatingtheseresultswithnovelspeakersandlanguages
shouldprovidefurtherevidencethattalkerlanguagebackgroundaffectsadaptation.Inthe
Non-nativeDifferentcondition,however,thewithinnon-nativedistinctionshypothesis
predictsahighdegreeofadaptation,withlistenersinferringthatdifferentnon-native
languageshavedifferentphonotacticgrammars.Thenativevs.non-nativehypothesis,on
theotherhand,predictsasimilar,moderatedegreeofadaptationintheNon-native
DifferentandNon-nativeSharedconditions,asunderthishypothesislistenersdon’t
distinguishbetweendifferentnon-nativephonotacticgrammars.
57
2.5.1. Participants
192participantswererequired(64participantsforeachofthe3conditions).To
reach192participantswhopassedtheexperimentcriteria(seebelow),441participants
wererecruited,202ofwhompassedthecriteria(45.8%).10participantswhopassedthe
criteriawereexposedtoanexperimentallistapreviousparticipanthadbeenexposedto
andassuchwereexcluded.AsinExperiment1,participantswererequiredtohaveaU.S.IP
address.98%ofparticipantsself-identifiedasnativespeakersofEnglish,while1
participantself-identifiedasspeakinganon-NorthAmericandialectofEnglish.98.4%of
participantself-identifiedashavingnospeechorlanguageimpairments.Allmodelresults
werequalitativelyidenticalwhennon-nativeandparticipantsandthosewithimpairments
wereexcluded.
2.5.2. Stimuli
Stimuliwererecordedinasoundproofboothata44.1kHzsamplingrateand
normalizedto60dBSPL.6talkersrecordedstimuli,with1maleand1femalespeakerfor3
languages:English,Hungarian,andHindi.Talkersproduceddisyllablesfromorthographic
representationsofdisyllablesonamonitor;orthographyreflectedthelanguage
backgroundofthespeaker(atransliteratedorthographywasusedforHindi).AllHindiand
Hungariantalkerswerebilingual,butwereinstructedtoproducestimuliasdisyllablesin
theirnativelanguage,ratherthanEnglish.Disyllableswerepresentedinarandomorder.
58
Giventheaddeddifficultyofdetectingdifferencesintalkerlanguagebackground
betweentalkersoftwonon-nativelanguages,stimuliconsistedofdisyllablesratherthan
monosyllablestoprovidelistenerswithgreaterphoneticevidenceoftalkerlanguage
background.ThesyllablesmakingupthedisyllabicstimuliinExperiment2wereasubset
ofthoseusedinExperiment1.Consonantsconsistedofvoicelessstops[p,k]andvoiceless
fricatives[f,ʃ];[t]and[s],whichcanformcomplexonsetsinthesecondsyllable,werenot
usedtoensurethateachindividualsyllablewasparsedasconsonant-vowel-consonant.For
EnglishandHindispeakers,vowelsconsistedof[i]and[u];forHungarianspeakers,vowels
consistedof[i]and[y].Thisresultedinatotalof32monosyllables(4onsets*2vowels*4
codas).
64disyllabicstimuliwerecreatedbysplittingthe32monosyllablesinto4groupsof
8,counterbalancedforcodapattern(fricativevs.stop)andonset/rhymepattern(onset
[k,f]matchedwithrhymes[uf,ih,uk,ip]vs.onset[h,p]matchedwithrhymes[if,uh,ik,
up]).Thesegroupsof8arefurthersplitintwo,suchthateachgrouphasaneven
distributionofsegmentsineachposition.Eachsubgroupof4iscrossedtocreate32
disyllables(4syllables*4syllables*2positions).Amongtheresulting128disyllables,all
disyllableswithgeminationandreduplicationareremoved,andsubsetswerechosensuch
thatsyllablesappearedanequalnumberoftimesinbothpositionswithineachgroup,fora
totalof64disyllables.
59
2.5.3. Procedure
TheprocedurewasidenticaltothatofExperiment1.
2.5.4. Design
Stimuliweresplitinhalf,intogeneralizationandfamiliarizationdisyllables(32
each),byonset-vowel/codapairings,andcounter-balancedacrossparticipants.For
example,indisyllablesthatincludesyllablesendingincodas[k,f],ParticipantAhearsthe
onset-vowelpair[fu_](e.g.,fookpeek)infamiliarizationdisyllables,andtheonset-vowel
pair[fi_]innovelgeneralizationdisyllables(e.g.,feekpook).ParticipantBhearsthe
conversepattern(e.g.,feekpookinfamiliarization;fookpeekingeneralization).Amongthe
32familiarizationdisyllables,syllablesinhalf(16)endinfricatives,whilesyllablesinthe
otherhalfendinstops.
Thesetsweresplitinhalfagainintosubsetsof8,suchthateachsyllableonly
appearsonceineachposition(e.g.,fifappearsonceasthefirstsyllableandonceasthe
secondsyllable).Todecreasetheoverallconfusabilityofthesets,participantsheareach
speakerproduceonlyonesubsetof8infamiliarization(althoughtwiceasoften;see
below),whiletheothermatchingsubsetiswithheld.AsinExperiment1,eachspeaker
repeatsfamiliarizationdisyllablesthatendinadifferentcodapattern(e.g.,SpeakerAends
theirsyllablesinstops;SpeakerBinfricatives);whichtalkerproduceswhichsetis
counterbalancedacrossparticipants.
60
Amongthe32generalizationsyllables,eachspeakerproduceshalf(16)oftheset.
Amongthissubset,half(16)followtheconstraintestablishedinthefamiliarizationset,and
halfviolatethisconstraint(i.e.,bothspeakerssaynovelgeneralizationdisyllablesthatend
inbothstopsandfricatives).
Thefirst2blocksoftheexperimentconsistsofthefamiliarizationphase.Ineach
block,participantsareexposedto2repetitionsofeachofthe16familiarizationdisyllables
(halfsaidbyeachspeaker)inrandomorder,foratotalof32tokensperblock.Pilottesting
suggestedthatduetotheincreasedsimilarityoftokensinthisstudy,tworepetitionsof
eachdisyllableperblockwererequiredtoensureadequatelevelsofrecognition
performanceonthefamiliarizationtokens.Inthegeneralizationphase,theserandomized
setsof32tokensarerepeatedin8furtherblocks;eachgeneralizationblockalsoincludes4
intermixedgeneralizationdisyllables.Thisresultsinatotalof352trials(16familiarization
disyllables*2repetitions/block*10blocks+32generalizationsyllables).
Toensurethatlistenerscanclearlytelltalkersapart,thetwotalkershavedifferent
gendersineachcondition.
2.5.5. DataAnalysis
AsinExperiment1,participantshadtopassasetofcriteriatoensurethatthey
adequatelyattendedtothetask.ToachievesimilaroverallpassingtothoseinExperiment
1,giventheincreasedconfusabilityofthefamiliarizationsetinExperiment2,thecriteria
61
forperformancewereslightlylowered:participantshadtocorrectlyacceptatleast85%of
familiaritems(asopposedto90%inExperiment1).AsinExperiment1,participantshad
tocorrectlyrejectatleast10%ofthenovelgeneralizationitemsthattheyhadnotheard.
Finally,athirdcriterionwasincluded:participantshadtocorrectlyrejectnovel
generalizationitemsatleastasoftenastheyrejectedfamiliarizationitems(e.g.,ifa
participantcorrectlyacceptedfamiliarizationitemsonly85%ofthetime,theparticipant
hadtocorrectlyrejectgeneralizationsyllablesatleast15%ofthetime).Thisensured
participantswhowereunabletosufficientlydifferentiatefamiliarandnovelitemswerenot
includedintheanalysis.
Generalizationdatawasanalyzedusingalogisticmixed-effectsregression.Fixed
effectsincludedlegalityandtwocontrast-codedterms:languagedifference,inwhichthe
Non-NativeSharedconditionwascontrastedwiththetwoDifferentconditions;andnon-
nativelanguagebackground,inwhichtheNon-nativeSharedandNon-nativeDifferent
conditionscontrastedwiththeMixedDifferentcondition.Furthermore,themodelincluded
aninteractiontermbetweenlegalityandeachofthecontrast-codedterms.Therandom
effectsstructureincludedrandominterceptsandrandomslopesoflegalitybyboth
participantsanditems.
Thewithinnon-nativehypothesispredictsasignificantdifferencebetweentheNon-
nativeSharedandthetwoDifferentconditions,asindicatedbytheinteractionterm
betweenlegalityandthenon-nativeterm;thenativevs.non-nativehypothesisdoesnot
predictsuchadifference.Suchadifferencewouldindicatethatlistenersshowedalarger
62
legalityadvantageintheDifferentconditions,despiteoneoftheseconditionsincluding
speakersoftwodifferentnon-nativelanguages.
2.5.6. Results
Participantscorrectlyaccepted89.3%offamiliarizationdisyllables(CI[88.7%,
90.0%])andfalselyrecognized69.8%ofgeneralization.Themeanlegalityadvantagewas
14.0%(CI[11.3%,16.5%]).Crucially,asshowninFigure2.3,thedifferenceinlanguage
backgroundmodulatesthelegalityadvantage:similartoExperiment1B,thelegality
advantageismoderateintheNon-NativeSharedcondition,andlargeinbothDifferent
conditions.
Figure2.3.A:FalserecognitionratesforlegalandillegalgeneralizationsyllablesinExperiment2.B:Legalityadvantage(falserecognitionrateonlegalgeneralizationsyllablesminusfalserecognitionrateonillegalgeneralizationsyllables)forExperiment2.Inbothpanels,errorbarsreflectbootstrapped95%confidenceinterval.
63
Theresultsfromthelogisticmixedeffectsregressionshowamaineffectoflegality(β=
0.76,s.e.β=0.08,χ2(1)=62.83,p<.0001),showingthatlistenersweremorelikelytofalse
alarmonlegaldisyllables.Theinteractionbetweenlegalityandlanguagedifferencewas
alsosignificant(β=0.71,s.e.β=0.21,χ2(1)=11.1,p<.001),aslistenersshowedagreater
legalityadvantageontheDifferentconditionsthanintheSharedcondition.Legalitydidnot
interactwiththenon-nativebackgroundterm(β=-0.04,s.e.β=0.21,χ2(1)=0.03,p=
0.87),aslistenersdidnotshowalargerlegalityadvantageinthetwoNon-nativeconditions
vs.theMixedcondition.
2.5.7. Discussion
ListenersinExperiment2adaptedtotalker-specificconstraintsineachcondition.
ThisreplicatesfindingsfromExperiments1AandBusingnoveltalkers,languages,and
stimulusdesign,providingfurtherevidencethatlistenerscanadapttotalker-specific
constraints.AsinExperiments1AandB,thedegreeofadaptationwasmodulatedbythe
languagebackgroundofthetalkers:listenersshowedahighdegreeofadaptationwhen
talkersdifferedinthelanguagebackground(MixedDifferentandNon-NativeDifferent
conditions),andalow-to-moderatedegreeofadaptationwhentalkerssharedalanguage
background(Non-NativeShared).
ListenersadaptedatasimilarrateinbothDifferentconditions,regardlessof
whethertheywereexposedtooneHindiandoneHungariantalker(Non-NativeDifferent)
64
oroneEnglishandoneHindi/Hungariantalker(MixedDifferent).Thissuggeststhat
listenersmakedistinctionsbetweendifferentnon-nativephonotacticgrammars,andassign
differentphonotacticgrammarstodifferentnon-nativelanguages.Inotherwords,ifthe
phoneticsoftwolanguagesareperceptiblydifferent—regardlessofwhethertheyare
nativeornon-nativelanguages—listenerscaninferthatthoselanguageshaveseparate
phonotacticgrammars.
HowcanweunderstandthesimilarlevelsofadaptationintheNon-NativeDifferent
andMixedDifferentconditions?Theseresultsareconsistentwiththehypothesisthat
participantsaretreatinginputtheyreceiveinphonotacticlearningexperimentsasanon-
native“lablanguage”.Participantsareexposedtononsensewordsinsemantically
meaninglesscontexts,andlikelyinfersuchlanguagesarenotnative.Thespeedwithwhich
participantsadapttonovelphonotacticconstraints(withinonetotwosessions)also
suggeststheyarelearninganovelphonotacticgrammar,ratherthanadjustingonewith
whichtheyhavealifetimeofexperience.Thisisfurthersupportedbyevidencein
productionthatspeakersmaintainexperimentalconstraintsforatleastaweek,and
possiblylonger,despiteinterveningevidencetothecontraryoutsideoftheexperiment
(Warker,2013).Inpreviousperceptionexperiments,listenerslearnrapidlydespitebeing
exposedtoasinglenativetalker.Inotherwords,itislikelythatlearnerstreatanyspeaker
intheexperimentalcontextasbeing“non-native”,evenifthatspeakerhasphonetics
consistentwithEnglish.ThustheMixedDifferentconditionhastwo“non-native”languages,
despiteoneofthemhavingEnglishphonetics.Assuchwedon’texpectdifferencesbetween
theMixedDifferentandNon-nativeDifferentconditions,givenlistenersmayinfertheyboth
65
havespeakersoftwodifferentnon-nativelanguages.Notethatthisiscompatiblewiththe
moderateadaptationfoundintheNon-nativeSharedconditioninExperiment1.Weargue
thatthisisduetotheasymmetryinknowledgeofnativevs.non-nativephonetics.Inother
words,listenersarelikelytoknowthattwonativespeakersofEnglishsharealanguage
background;theyarelesslikelytoknowthattwoFrenchspeakersdo.
WhileExperiment2replicatedtherelativelyhigherlegalityadvantageinDifferent
vs.SharedconditionsfoundinExperiment1,theoveralllegalityadvantagesarelowerin
Experiment2(e.g.,inExperiment1BthemeanlegalityadvantageintheDifferent
conditionsis22.5%;inExperiment2it’s17.5%).Totheextentthatthesedifferencesin
effectsizesbetweenexperimentsaremeaningful,itislikelyduetodifferencesinthe
designsofthetwoexperiments.InExperiment2,thestimulussetwasmuchmore
confusablethaninExperiment1.Thislikelycausedtherelativelyhighoverallfalse
recognitionrate(57.7%inExperiment1;69.8%inExperiment2).Thisalsomayhave
loweredthelegalityadvantage,asparticipantsmayhavebeguntohitaceilingonfalse
recognitionratesforlegalsyllables.
2.6. ListenerLanguageBackgroundAnalysis
ResultsfromExperiments1and2stronglysuggestthatpreviousexperiencewith
non-nativelanguages,andthephonotacticvariationthatdifferentlanguagesexhibit,
constrainlistener’sadaptationtonovelnon-nativephonotactics.Listenersinferthat
speakersofdifferentnon-nativelanguagesareunlikelytoshareaphonotacticgrammar,as
66
opposedtospeakerssharinganon-nativelanguagebackground.Buthowmuchexperience
withnon-nativelanguagesisnecessarytomakesuchinferences?It’spossiblethatthe
thresholdisquitelow,withmonolingualspeakersabletomakesuchinferencesthrough
theirdailyexposuretonon-nativelanguages.Alternatively,multilingualspeakersmay
morereadilymaketheseinferencesbasedontheirpastexperiencelearninglanguages.
Participantsreportedtheirlanguagebackgroundsonaquestionnairebeforetaking
theexperiment.Participantsreportedthelanguagestheyknow,theirageofacquisition,
andthelengthoftimespeakingthoselanguages.Anyparticipantswithanageof
acquisitionof5yearsoldorearlierwereclassifiedashavingearlysecondlanguage(L2)
experience.33.8%ofparticipants(216total)reportedspeakingatleastonelanguageother
thanEnglish,while7.7%(49)hadearlyL2experience.AmongparticipantsinExperiment
1,inwhichparticipantswereexposedtoFrenchspeakers,6.7%(30)reportedknowing
someamountofFrench.AmongtheparticipantsinExperiment2,whowereexposedto
HindiandHungarianspeakers,nonereportedknowingHindiorHungarian.
Overall,participantswhohadanyL2experienceshowedameanlegalityadvantage
(i.e.,falserecognitionratesforlegalstimuliminusfalserecognitionratesforillegalstimuli)
of15.8%(CI[13.6%,18.1%]),andanoverallfalserecognitionrate,regardlessoflegality,of
61.0%(CI[58.2%,63.6%]);participantswithnoL2experienceshowedalegality
advantageof14.7%(CI[13.1%,16.4%])andafalserecognitionrateof61.5%(CI[59.5%,
63.6%]).ParticipantswithearlyL2experience,ontheotherhand,showedan18.9%
legalityadvantage(CI[14.6%,23.0%])and57.3%falserecognitionrate(CI[51.5%,
62.7%]);participantswithlateornoL2experienceshoweda14.8%legalityadvantage(CI
67
[13.4%,16.2%])and61.7%falserecognitionrate(CI[60.0%,63.3%]).Listenerswith
FrenchexperienceinExperiments1AandBshoweda13.1%legalityadvantage(CI[6.3%,
16.5%]),whilelistenerswithnoFrenchexperienceinExperiments1AandBshoweda
15.6%legalityadvantage(CI[13.1%,16.6%]).
Mixed-effectsregressionswereusedtoassessdifferencesbasedonL2experience.
SeparatemodelswererunforearlyL2experience,anyL2experience,andFrenchL2
experience.ForearlyL2andanyL2experience,datawaspooledoverbothexperiments;
forFrenchL2experience,onlydatafromExperiment1wasincluded.Thesemodelsdidnot
showaneffectofL2experienceonthelegalityadvantage,astherewerenosignificant
interactionsbetweenthelegalityandL2experienceterms.Moreover,therewereno
significantmaineffects(i.e.,overalldifferencesinfalserecognitionrates)basedonL2
experience.
Basedontheseresults,itdoesnotappearthatparticipantswithL2experiencemade
strongerinferencesaboutthephonotacticgrammarsofspeakersbasedonthespeaker’s
languagebackgrounds.Thissuggeststhatarelativelysmalldegreeofexposuretonon-
nativephonotacticsisrequiredtomakeinferencesabouttalker’sphonotacticgrammars
basedontheirlanguagebackground,aslistenerswithoutextensiveL2experiencedoso.
Thisisnotsurprisinggivenlisteners’sensitivitytonon-nativephonotactics,evenininfants
asyoungas9-months-old(MattysandJusczyk,2001).Listenersalsotakeintoaccount
talkerphonotacticswhenjudgingspeakeraccentedness.Whenlistenershearspeechinthe
speaker’sL2,sequencesthatarelegalinthespeakers’nativelanguage(L1)aredeemedless
accentedthansequencesillegalinthespeakers’L1(Park,2013).Inotherwords,
68
monolinguallistenersarehighlysensitivetonon-nativephonotactics,andarelikelyto
attendtosuchnon-nativepatternswhentheyappearintheinput,evenifthatinputis
relativelylimited.Thatsaid,therewereanumberoflimitationstothecurrentanalysisof
listenerlanguagebackground:first,listenersself-reportedtheirownlanguagebackground,
andwerenotaskedtoassesstheirownproficiency(althoughfluencyself-assessmentsmay
notbeentirelyreliable;Tomoschuk,Ferreira,&Gollan,2018).Thismeanstherewaslikely
awiderangeofproficiencylevelswithinlistenerswithL2experience.Second,therewasa
relativelysmallnumberoflistenerswithL2experience(especiallyearlyL2experience),
whichmayhaveresultedintoosmallofasampletodrawdefinitiveconclusions.Future
workshouldinvestigatetherelationshipbetweenlistenerlanguagebackgroundand
inferencesaboutphonotacticvariationdirectly,bycomparingbilingualandmonolingual
populations.
2.7. GeneralDiscussion
Recentresultsfromtheperceptualadaptationliterature(e.g.,LiuandJaeger,2018)
suggestthatlistenersusetheirpastexperiencetouncovertheunderlyingstructurethat
generatesvariationinspeechforms,andmakecausalinferencesbasedonthisstructure
whenexposedtonovelinput.Inthecaseofphonotactics,thereismassivevariation
betweenthephonotacticsystemsofdistinctlanguages,andrelativelylittlevariationwithin
asingledialect.Ourpredictionthatlistenerswouldleveragethispastexperiencewith
phonotacticvariationtomakeinferencesaboutnovelphonotacticconstraintsduring
69
adaptationwaslargelyconfirmedinthreeexperiments.Experiments1AandBfoundahigh
degreeofadaptationtonovelphonotacticconstraintsbyEnglishlistenerswhentalkers
differedinlanguagebackground(Frenchvs.English).Incontrast,adaptationwasmoderate
whentalkerssharedanon-nativelanguagebackground(twoFrenchtalkers)orrelatively
lowwhentalkerssharedanativelanguagebackground(twoEnglishtalkers).Experiment2
showedthatthehighdegreeofadaptationincaseswheretwospeakersdifferintheir
languagebackgroundgeneralizestonon-nativelanguages(Hindivs.Hungarian)aswell.
Thispatternofresultssupportsthehypothesisthatlistenersmakedistinctionsbetween
non-nativephonotacticgrammars,andusethisinformationwhenmakinginferencesabout
whetherornottalkerssharedaphonotacticgrammar,althoughfurtherreplicationis
necessarytoconfirmthatthisisreliableacrossarangeoftalkers.
Whilelearningwasstrongerwhentalkersdifferedinlanguagebackground,learning
alsooccurredwhentalkerssharedalanguagebackground.Wearguethatthemore
confidentlistenersarethattwotalkersdifferintheirlanguagebackground,thegreater
adaptationwillbe;conversely,whenlistenersareconfidentthatthetalkerssharea
languagebackground,adaptationwillbelower.Whenlistenersareexposedtotwonative
Englishtalkers,forexample,wepredictedtheywouldbehighlyconfidentthatthetwo
talkerssharealanguagebackground(andthereforeaphonotacticgrammar),andwould
notadapt.Surprisingly,listenerswereabletoadaptinthiscondition,incontrastto
previousworkreportingnullresults(Onishietal.,2002).Listeneradaptationwaslowestin
thiscondition,however,suggestingthatlistenershadastrongpriorbeliefthatthetwo
Englishtalkerswouldhavesimilarphonotacticsrelativetootherpairsofspeakers.It
70
appearstheevidencethatthetwotalkerssharedalanguagebackground,intheformof
theirsimilarphonetics,wasn’tenoughtoentirelyovercometheevidencetothecontrary.
Thisincludesthedifferentphonotacticpatternsexhibitedbyeachtalker;thefactthat
listenerswerelisteningtonon-nativewords;andthatasingleparticipantneverheardboth
talkersproducethesameword.
Theeffectofconfidenceconcerningtalkerlanguagebackgroundismoreacutewhen
twonon-nativetalkerssharethesamelanguagebackground,astheydidintheNon-Native
Sharedcondition.Here,wefoundagreaterdegreeofadaptationrelativetotheNative
Sharedcondition.Thisislikelyduetotheasymmetrybetweenlistenerknowledgeofnon-
nativevs.nativephonetics.NativeEnglishlistenershavelessknowledgeoftheFrench
phoneticsystemthantheEnglishone;thereforewhilethelistenersmayperceivetwo
Frenchspeakersasphoneticallysimilar,listenerswon’tbeasconfidentastheyarefortwo
Englishspeakers.
2.7.1. PhonotacticsandL2Acquisition
Whyisphonotacticadaptationsorapid,robust,andflexible?Intheseexperiments,
listenerswereabletosimultaneouslyadapttotwodistinct,complex(i.e.,second-order)
phonotacticconstraintswithinasingleshortexperimentalsession,showingsensitivityto
differentnon-nativelanguagesandevenindividualspeakers.Onepossibilityisthat
71
phonotacticsareacriticaltoolintheearlieststagesofL2acquisition.7Thismaybebecause
phonotacticconstraintsguidespeechperceptionbylimitingthenumberoflexicaland
phonologicalcandidateslistenershavetoconsider(e.g.,listenersperceiveambiguous
soundsastheoptionthatresultsinalegal,ratherthanillegal,sequence;Massaroand
Cohen,1983).Thismaybeparticularlyimportantwhenspeechperceptionislessaccurate
intheearlystagesofacquisition.Phonotacticsalsoactasanimportantcueinword
segmentation(McQueen,1998),whichinturnisaprecursortolexicalacquisition.Indeed,
someevidencesuggeststhatadultlistenerslearnnovelL2wordswithhighphonotactic
probabilitymoreeasilythanthosewithlowprobability(Storkel,Armbrüster,&Hogan,
2010).
Forthelearner,adaptingtosubsetphonotacticsappearstobefairlyeasyrelativeto
othertypesofadaptation(e.g.,acquiringperceptualdistinctionsbetweentwoL2sound
categoriesthatassimilateintoasingleL1category;e.g.,Best,McRoberts,andGoodell,
2001).Inotherwords,whenthelistenerisfacedwiththeoverwhelmingprospectof
acquiringandunderstandinganunfamiliarlanguage,adaptingtothelanguage’ssubset
phonotacticsmayserveasacognitivelyinexpensiveadjustmentthataidslistenerswith
someofthemostimportantearlytasksofacquisition:speechperception,word
segmentation,andlexicalacquisition.Ifthisisthecase,wewouldpredictthatlearnerswho
7Thismaybeparticularlytrueforthetypeofsubsetphonotacticconstraintslistenerswereexposedtointheseexperiments(i.e.,restrictionsonsequencesthatarelegalinthelistener’sL1),whichdonotpresentthesameperceptualissuesassupersetphonotactics(i.e.,difficulttoperceivesequencesthatareillegalinthelistener’sL1,suchas[dl]inwordonsetforEnglishlisteners;seeDupoux,etal.,1999).
72
areabletomoresuccessfullyadaptinphonotacticadaptationexperimentswouldalsobe
moresuccessfulinearlyL2acquisition.
It’salsopossibletheseadjustments,atleastinperception,areshort-lived,which
mayexplainthespeedwithwhichtheyaremade:earlyinacquisition,learnersmayneedto
readjustwitheachexposuretothenon-nativelanguage.Finally,thereisthequestionof
whylistenerscanlearnmultiplenon-nativephonotacticgrammarssoquickly.Ofcourse,we
knowthismustbepossibleinthelongterm,sincemultilingualspeakershaveaccessto
threeormoredistinctphonotacticgrammars.Furthermore,thismaybeaparticularly
importantskillincontextswithlanguagecontactoccurringbetweenmultiplelanguages.
2.7.2. AccentDetection
Listenersarecapableofmakingremarkablyfine-graineddistinctionsinnon-native
accentdetection(AtagiandBent,2013).However,incaseswherelistenersareexposedto
singlewordswithsubsetphonotactics,asinthecurrentexperiments,thisisahardertask
(Park,2013).Howwerelistenersabletodothis,particularlyinExperiment2,inwhich
theyweretaskedwithdistinguishingbetweentwonon-nativeaccents?Onepossibilityis
thetypeofcomparisonlistenershadtomakeinthecurrentexperimentswaseasierthan
thoseinpreviousstudies.Inmanyaccentdetectionstudies,listenersaregivenanauditory
freeclassificationtask,inwhichtheycomparealargenumberofdifferentaccented
speakersatonceonagradienttwo-dimensionalscale(e.g.,AtagiandBent,2013).Whilethe
highnumberofdifferentaccentsincreasesthedifficultyofthistask,listenersareallowed
73
totakeaslongastheywantandre-listentoaudiosamplesofeachspeaker.Otherstudies
useatwo-alternativeforcedchoicediscriminationtask,inwhichlistenersarepresented
withanativeandnon-nativespeakerproducingthesameitemandmustdistinguish
betweenthem(e.g.,Park,2013).Inthecurrentstudy,thelistener’staskislessdemanding
thanthesepreviousstudies:listenersonlymust(implicitly)decidewhethertwospeakers
havedifferentaccents,notwhichoneisthenativevs.non-nativespeaker.Moreover,only2
languagesatatimeareinvolved.Finally,listenersdonothavetodecidewitheachitem
whichspeakerisnativeandwhichisnon-native;theyareabletobuildtheirrepresentation
ofthespeaker’slanguagebackgroundoverthecourseoftheexperiment.
2.8. Conclusion
Inthreeexperiments,wehaveshownthatlistenersusetheirpriorexperiencewith
phonotacticvariation—thatlanguagevaryintheirphonotacticsmuchmorethanindividual
speakersofthesamedialect—toguidetheiradaptationtonovelphonotacticconstraints.
Listenersevaluatetheunderlyingstructuregeneratingphonotacticvariation,andexhibita
largedegreeofadaptationtosystematicsourcesofphonotacticvariation(i.e.,listenerswho
differintheirlanguagebackground),andasmallerdegreeofadaptiontoincidentalsources
ofvariation(i.e.,listenerswhosharealanguagebackground).Thiseffectextendsto
differencesbetweendifferentnon-nativelanguages.Together,theseresultsilluminatea
corelinguisticability:appropriatelyadaptingtoourdynamiclanguageenvironmentbased
onourpriorexperience.
74
3. Study2
3.1. Introduction
Speechisproducedwithahugeamountofvariationbetweentalkersandcontexts.
Toeffectivelycomprehendtheirinterlocutors,speakersmustcontinuallyadapttonovel
speakersandcontexts,drawingontheirexperiencefromsimilarsituationsinthepast.
Adaptationisnotlimitedtocomprehension,however—communicationisatwo-waystreet,
withspeakerstradingoffproducingandcomprehendingspeech.Speakersnotonlymodify
theirpredictionsofwhattheyexpecttohear;theyalsomodifytheirownspeech,imitating
andaligningthemselveswiththephoneticcharacteristicsoftheirinterlocutorin
spontaneousspeech(e.g.,Pardo,2006)orwhileshadowingamodeltalker(e.g.,Goldinger,
1998).
Phoneticimitationislikelymotivatedbyacomplexmixtureofsocialand
communicativefactors(Pardo,Urmanche,Wilman,&Wiener,2017),butitisonlymade
possiblebythesystematicityofinter-speakervariation.Imitatinganinterlocutor—either
tobringoneselfinclosersocialalignmentwithaninterlocutorinapositionofpower(e.g.,
Giles,1973)ortosimplifylanguageprocessingforbothspeakersinadialogue(e.g.,
PickeringandGarrod,2004)—onlyhasutilitybecauseindividualspeakersvary
systematically,showingconsistencyfromoneutterancetothenext.Ifspeakersvaried
freely,suchimitationwouldbringyounoclosertoyourinterlocutor’sspeech.
Inadditiontospeakersadaptingtothephoneticfeaturesofinterlocutorsandmodel
talkers,inexperimentalsettingsspeakersalsoadapttheirspeechproductionsystemsto
75
reflectnovelphonotacticconstraintsembeddedinlaboratoryspeechortext(e.g.,Delletal.,
2000;Onishi,Chambers,&Fisher,2002).Theseexperimentalconstraintsareoften
arbitrary(e.g.,[n]isconstrainedtoonsetposition;[f]isconstrainedtocodaposition)and
are,bydefinition,notcharacteristicofaspeaker’snativelanguage.Intonguetwister
paradigms,forexample,participants’speecherrorsreflectthephonotacticconstraintsof
thelaboratorytexttheyareexposedto(e.g.,errorsresultingin[n]rarelyappearincoda
position).
Inthisstudy,weexplorehowthedifferencesinthestructureofvariationfor
phoneticsandphonotacticsmayresultindifferencesinadaptation.Whileadaptationtothe
phoneticpropertiesofaninterlocutorismotivatedbyspeakers’priorexperiencewiththe
systematicnatureofinter-speakervariation,phonotacticconstraintsvarylittleatthe
individuallevel.Instead,phonotacticsvaryextensivelyatthelanguage-anddialect-wide
level,withdifferentlanguagevarietiesexhibitingabroadrangeofpossiblesyllable
structuresandsoundsequencesnotfoundwithinlanguagevarieties.
Weexpectthisdifferenceinthestructureofvariationforphonotacticsversusthat
forphoneticstoresultindifferencesbetweenphoneticandphonotacticadaptation,aswe
sawinspeechperceptioninStudy1.Wearguethatspeakersmakecausalinferencesabout
thesourceofphonotacticvariationwhenadapting(e.g.,LiuandJaeger,2018),guidedby
theirpriorexperiencewithvariation:thattalkerswithinalanguagevarietyshareasingle
phonotacticgrammar,whiletalkerswhodifferintheirlanguagevarietiesdonot.Assuch,
wepredictthatwhenspeakersencountertwotalkerswithdifferingphonotactics,theywill
onlyadapttoeachtalker’sphonotacticgrammariftheybelievethosetalkersdonotsharea
76
languagebackground.ThishypothesisisconsistentwithOnishietal.(2002),whofound
thatinaspeededrepetitiontask,participantswereabletolearnsecond-orderconsonant
constraintsconditionedonsurroundingvowels,butnotconstraintsconditionedontalker
identity(bothtalkerswerenativeEnglishspeakers).
Weexaminethisquestionusingamodifiedtonguetwisterparadigm.Oneachtrial,
participantswillfirstshadowamodelspeakerproducingastringoffoursyllables,andthen
repeattheirproductionsagainwithorthographicsupport(previoustonguetwister
experimentshavepresentedthetwisterinorthographicformonly,withtheexceptionof
Smalle,Muylle,Szmalec,&Duyck,2017).Thestringsofsyllablesassociatedwitheach
modeltalkerwillreflectdifferentphonotacticconstraints.Thelanguagebackgroundofthe
modeltalkerswasmodulatedindifferentconditions,withmodeltalkerseithernative
speakersofGermanorEnglish.ThreeofthefourconditionsmirrortheconditionsinStudy
1,withtwotalkersthateithershareordifferintheirlanguagebackgrounds:NativeShared
(twomonolingualEnglishmodeltalkers),Non-NativeShared(twonativeGermanmodel
talkers),andDifferent(oneGerman,oneEnglishtalker).Inaddition,aVowelconditionwas
includedasataskcontrol,ensuringthatresultsfrompreviousworkarereplicated(e.g.,
Gaskell,Warker,Lindsay,Frost,Guest,Snowdon,&Stackhouse,2014;Smalleetal.,2017;
Warker,2013;Warker&Dell,2006)despitechangestothetask.IntheVowelcondition,
participantsareexposedtoasinglemodeltalker,andlearnasecond-orderconstraintin
whichconsonantrestrictionsareconditionedontheneighboringvowel(e.g.,[ɛ]canbe
followedby[m]andprecededby[n];thereverseistruefor[ɪ]).Adaptationismeasuredby
analyzingtheerrorsofparticipants:iferrorsinvolvinganexperimentallyconstrained
77
segmentresultinlegalsyllables(i.e.,syllablesthatfollowthephonotacticpattern
embeddedintheinput)morethanillegalsyllables,itsuggestsparticipantsareadapting.
Ifthesameprinciplesforadaptationapplyinproductionastheydoinperception,
weexpectasimilarresulttowhatwasfoundinpreviousperceptionexperiments:ahigher
degreeofadaptationintheDifferentconditionthaninthesharedconditions.Atamore
granularlevel,wemayexpectmoderateadaptationintheNon-NativeSharedcondition,and
alowdegreeofadaptationintheNativeSharedcondition,mirroringtheresultsfrom
perception.
3.2. Background
3.2.1. Phonotacticlearninginspeechproduction
Overthepasttwentyyears,researchershaveexploredphonotacticadaptationusing
thetonguetwisterparadigm(Anderson,Holmes,Dell,&Middleton,2019;Delletal.,2000;
Gaskelletal.,2014;Goldrick,2004;Goldrick&Larson,2008;Kittredge&Dell,2016;Smalle
etal.,2017;Taylor&Houghton,2005;Warker,2013;Warker&Dell,2006;Warkeretal.,
2008;Warker,Xu,Dell,&Fisher,2009).Inthetonguetwisterparadigm,errorsareelicited
byhavingparticipantsquicklyrepeatastringofnonsensesyllables.Thesesyllablesfollow
anartificial,experimentalphonotacticconstraint,suchas“syllablesbegin,butdonotend,
in[n];viceversafor[f]”.Inaddition,twistersalsoincludesegmentsthatareconstrainedin
theparticipant’snativelanguage,suchas[ŋ],whichisconstrainedtocodapositionin
English.Speecherrorsvirtuallyneverviolatesuchlanguage-widecategoricalphonotactic
78
constraints—thisissometimesreferredtoasthephonotacticregularityeffect(Fromkin,
1971).InEnglish,forexample,producingfipwhenthetargetissipisapossibleerror,but
ngipforsipisanextremelyunlikelyerror.Otherconsonantsintheexperimentare
unconstrained,appearinginbothonsetandcodaposition.Suchunconstrainederrorstend
tomaintaintheirsyllableposition(Nooteboom,1969):forexample,fikfip,inwhich[f]
maintainsitspositionintheonsetofthesyllable,isamorelikelyerrorforthetarget
sequencefiksipthanfiksif,inwhich[f]switchestothecodaposition.Intonguetwister
experiments,thissyllablepositioneffect—whichholdsinroughlythree-quartersof
errors—servesasabaselineforexperimentallearningeffects.Learningismeasuredasthe
degreetowhicherrorsinvolvingexperimentallyconstrainedconsonantsmaintaintheir
syllablepositionaboveandbeyondthesyllablepositioneffectforunconstrained
consonants.Inpreviousexperiments,adaptationtothenovelconstraintcauseserror
patternsforexperimentallyconstrainedconsonantstoresembleerrorpatternsoflanguage
constrainedsyllables,oftenmaintainingtheirsyllableposition95%ofthetimeormore
(e.g.,Delletal.,2000).
Phonotacticadaptationexperimentshaveshownthatspeakersarealsosensitiveto
first-versussecond-orderconstraints,witheachtypeofconstraintshowingdistinct
learningpatterns.First-orderconstraintsaredependentonlyonsyllableposition(e.g.,“[n]
isconstrainedtoonsetposition;viceversafor[f]”);second-orderconstraintsare
dependentontwofactors(e.g.,“ifthevowelis[ɪ],[n]isconstrainedtoonsetand[f]is
constrainedtocoda;ifthevowelis[ɛ],viceversa”).Suchsecond-orderconstraints
naturallyoccurwithintheworld’slanguages.InEnglish,forexample,[b]appearsinonset
79
positionprecedingthevowel[i](asinbeet),butisunattestedincodapositionfollowing[i]
(e.g.,teeb;KesslerandTreiman,1997).Previousworkhasshownthatsecond-order
constraintsarenotlearnedasthoroughlyasfirst-orderconstraints,witherrorslesslikely
tomaintaintheirsyllableposition.Inaddition,learningeffectsforsecond-orderconstraints
donotappearduringthefirstexperimentalsession(seeWarker,2013).Morespecifically,
sleepconsolidationappearstoplayacriticalroleintheacquisitionofsuchconstraints,as
participantsrequiresleepininterveningperiodsbetweensessionstoshoweffectsof
learning(Gaskelletal.,2014;Warker,2013).Thecurrentstudyexposesparticipantsto
second-orderconstraints,withconsonantpositioneitherdependentonvowelormodel
talkerlanguagebackground,dependingonthecondition.Assuch,weincludethree
experimentalsessions,eachondifferentdays,toensureparticipantsareabletoconsolidate
theconstraints.
3.2.2. Causalinferenceinphonotacticadaptationinproduction
Wehypothesizethatadaptationisspurredincontextsinwhichspeakersare
interactingwithtalkersthathavedifferentlanguagebackgrounds.Ifthisisthecase,why
haveparticipantsinprevioustonguetwisterexperimentsadaptedatall,whenthey’re
simplyreadingtwistersaloudandnotinteractingorshadowingotherspeakers?Asnoted
inStudy1,Warker(2013)foundthatparticipantsmaintaintheexperimentallylearned
phonotacticconstraintsaslongasoneweekafterinitialexposure,despitethemassive
interveningexperiencewithEnglishcontrarytosuchconstraints.Participantsalsolearn
80
thenewconstraintsrapidly,despitealifetimeofexperiencesuggestingsuchconstraints
arenotapartoftheirnativelanguage.Warker(2013)proposesamechanisticaccountfor
suchlearning,inwhichparticipantscreateacopyoftheirgeneralphonotacticgrammarto
useinexperimentalsettings.Intheexperiment,thecopyisupdatedtoreflectonlythose
featuresoftheexperimentalcontextthatarenottrueofthegeneralphonotacticgrammar.
Aremainingquestioniswhyspeakerscreatethecopyinthefirstplace—whydo
speakershavetheadaptiveabilitytoquicklymodifytheirphonotacticsinthisway?Of
course,speakersdonotregularlyfindthemselvesinpsycholinguisticexperiments,sothis
doesnotseemlikeapossiblemotivationforthisadaptiveability;moreover,speakers’
nativelanguagephonotacticsarelikelyrelativelystable(Pierrehumbert,2001),suggesting
phonotacticadaptationisnotdrivenbydifferencesbetweenspeakerswithinone’snative
language.Wearguethatintonguetwisterexperiments,asintheperceptualadaptation
exploredinStudy1,participantsarerecruitingtheirsecond-language(L2)learning
facultiestoadapttonon-nativephonotacticconstraints.Inthehighlyartificiallaboratory
setting,inwhichparticipantsareexposedtononsensesyllablesthatarenotpresentedina
semanticallymeaningfulnativelanguagesetting,participantsmaydetectthattheyareina
non-nativelanguagecontext.Detectingsuchchangesincontextisacriticalaspectof
learninginamulti-contextenvironment(Qian,Jaeger,&Aslin,2012).Moreover,wehave
strongevidencethatlearnersareabletodetectdifferentphonotacticenvironmentsand
constructmultiplephonotacticgrammarsintheformofmultilingualspeakers.Multilingual
learnersareexposedtoconflictingphonotacticpatternsinacquisition,andmustseparate
evidencefordifferentconstraintsbasedonthelanguagecontextinwhichtheyappear
81
(althoughbilinguals’phonotacticgrammarsarenotentirelyseparate;seeCarlson,
Blasingame,Goldrick,&Fink,2016).Learnerslikelydothisbyleveraginganumberof
differentcues,suchasdifferinglexicalitems,phonetics,prosody,talkervoices,andother
factors(e.g.,Bosch&Sebastián-Gallés,2001;Weiss,Gerfen,&Mitchell,2009).Inthisstudy,
wepredictparticipantswillinfermultiplelinguisticcontextsbasedonthephonetic
differencesbetweenthenativeandnon-nativemodeltalkers.
3.2.3. CurrentStudy
Recentevidenceintonguetwisterstudiessuggeststhatspeakersarerelatively
insensitivetohigh-levelinferencesaboutthesourcesofvariation,however.Delletal.
(2000)informedsomeparticipantsoftheexperimentalconstraintstheywouldbeexposed
to,andfoundlearningwasunaffectedbypriorknowledgeoftheconstraint(seealsoSmalle
etal.,2017;Warker&Dell,2006).Andersonetal.(2019),inatonguetwisterparadigm,
exposedlearnerstoafirst-orderconstraintandthenreversedthephonotacticconstraint
partwaythroughtheexperiment,beforereversingitbackagaininthefinalblock.If
participantsweresensitivetothechangeincontext(i.e.,thereversalofthephonotactic
constraint)andthesourceofthevariation,theyshouldhavebeenabletorapidlyreverse
theconstraint,andlearnthereversedconstraintfasterthantheoriginalconstraint,given
thatthereversedphonotacticinvolvedthesametargetconsonants,simplyinflipped
syllabicposition.Instead,participantslearnedthereverseconstraintmoreslowlythanthe
original,suggestingthatphonotacticlearningisincremental,andresistanttocausal
82
inferencesaboutchangesincontext.Andersonandcolleaguesarguethatphonotactic
adaptationisinternaltotheaspectoftheproductionsystemthatconstructssyllables,and
“notverycognitivelypenetrable”fromoutsidesystems.Moreover,thelearningsystemmay
bedomaingeneral:participantsshowremarkablysimilarpatternsfor“phonotactic”
learninginbutton-pushingtasks,inwhichdifferentfingersareassigneddifferent
“consonants”(Andersonetal.,2019;Rebei,Anderson,&Dell,2019).Thisfurthersuggests
thatlearnersmaybeinsensitivetolanguage-specificinferences,suchasthosebasedon
accent.Finally,adaptationintheproductionsystemmaysimplybemorecognitivelycostly
thanadaptationinperception,andthereforeslowerandmoreconstrained(Samuel,2011).
Anaccountinwhichspeakersareinsensitivetotop-downinferencesaboutthesourceof
variation,motivatedbytheseprevioustonguetwisterexperiments,wouldpredictlearners
areentirelyunabletodivideinputbetweenmodeltalkersandformmultiplephonotactic
grammars.ThisaccountthereforepredictsadaptationinthebaselineVowelconditionin
thecurrentstudy.AdaptationintheDifferentcondition,however,wouldsuggestthat
inferencesbasedonlanguagecontext,specifically,mayhaveaprivilegedstatusoverother
lessrelevanttypesofinformation.
Themeasureofinterestintonguetwisterexperimentsistherateatwhicherrors
maintainsyllableposition.Ifparticipantsareadaptingtotheexperimentalconstraint,they
shouldbemorelikelytomakeerrorsthatmaintainsyllablepositionforexperimentally
restrictedconsonants(i.e.,errorsthatfollowthephonotacticconstraint)thanfor
unrestrictedconsonants.Adaptationwillthereforebeindicatedbyahigherproportionof
errorsmaintainingtheirsyllablepositionforconstrainederrorsthanunconstrainederrors.
83
Forsecond-orderconstraints,adaptationshouldspecificallyappearinthesecondandthird
experimentalsessions,followingpreviousresults(e.g.,WarkerandDell,2006).
WepredictahighdegreeofadaptationinthebaselineVowelcondition,following
resultsfrompreviousexperiments.IfwedonotreplicatepreviousstudiesintheVowel
condition,itmayindicateissueswiththecurrentdesign(especiallyhavingtwisters
presentedauditorilyaswellasorthographically).Wealsopredictahighdegreeof
adaptationintheDifferentcondition,inwhichmodeltalkersdifferintheirnativelanguage
backgrounds,reflectingspeakers’useoftheirpriorexperiencewithphonotacticvariation
toadapttonovelconstraints.BothSharedconditions,ontheotherhand,shouldshowalow
tomoderatedegreeofadaptation,giventhatbothmodeltalkerssharealanguage
background.Morespecifically,wemayfindlowadaptationintheSharedNativecondition,
andmoderateadaptationintheSharedNon-Nativecondition,aswedidinperceptionin
Study1.Suchintermediateadaptationmaybeduetospeakers’asymmetricknowledgeof
nativevs.non-nativelanguages—speakersarelikelyquitecertainwhentheyareexposed
totwonativespeakersoftheirnativelanguagethatthetwospeakerssharealanguage
background,butlikelymuchlesscertainwhentheyareexposedtotwonon-native
speakers.
84
Table3.1.Summaryofconditions,alongwithmodeltalkerlanguagebackground,vowels,gender,anddegreeofadaptation.
NativeShared Non-NativeShared
Different Vowel
Modeltalkers 2English 2German 1English,1German
1Englishor1German
Vowels [ɛ] [ɛ] [ɛ] [ɛ]and[ɪ]PredictedDegreeofAdaptation
Low Low-to-moderate
High High
3.3. Methods
Thedesignandanalysisoftheexperiment—includingpredictions,numberof
participants,stimulusdesign,andmodelstructure—weredefinedbeforedatacollection,
unlessotherwisenoted,inapre-registrationontheOpenScienceFoundationplatform
(osf.io/uryc5/).
3.3.1. Participants
Sixteenparticipantswererecruited,allnativeAmericanEnglishspeakersfromthe
NorthwesternUniversitycommunitywithnospeechorhearingimpairments.The
experimentconsistedof3one-hoursessions,eachonadifferentday.Eachsessiontook
placenofurtherthanaweekapart.Participantswerepaid$40:$10foreachofthefirsttwo
sessions,and$20forthefinalsession,toincentivizeparticipantstoattendallthree
sessions.Participantswererandomlyassignedtodifferentconditions.
Beforetheexperimentbegan,participantscompletedaself-reportedlanguage
85
backgroundquestionnaire,reportinganysecondorthirdlanguages,aswellastheirageof
acquisitionandlengthofspeakingforeachlanguage.
3.3.2. Materials
Participantswereexposedtosecond-orderconstraintsinallconditions:inthe
NativeShared,Non-NativeShared,andDifferentconditions,constraintswereconditionon
modeltalkeridentity(e.g.,ifTalkerAisthemodeltalker,[m]isconstrainedtocoda
positionand[n]isconstrainedtoonsetposition;theconverseistrueforTalkerB).Inthe
Vowelcondition,participantswereexposedtoasecond-orderconstraintsimilartopast
experiments,inwhichthepossiblepositionsofexperimentallyconstrainedconsonantsare
conditionedonthevowel.Fourmodeltalkerswereincluded:onemaleandonefemale
Germanspeaker,andonemaleandonefemaleEnglishspeaker.IntheVowelcondition,
participantswereexposedtoasinglemodeltalker;adifferentmodeltalkerwasassignedto
eachparticipant.IntheNativeSharedcondition,participantswereexposedtobothEnglish
modeltalkers;intheNon-NativeSharedcondition,halfoftheparticipantswereexposedto
bothGermanmodeltalkers.IntheDifferentcondition,participantswereexposedtoone
GermanmodeltalkerandoneEnglishmodeltalker,whodidnotshareagender(e.g.,male
GermantalkerandmaleEnglishtalker).
Followingpreviousstudies,stimuliconsistedofCVCnonsensesyllablesmadeup
fromtwovowels([ɪ]and[ɛ])andeightconsonants([f],[p],[k],[t],[m],[n],[ŋ],[h]).While
mostpreviousexperimentsused[g]and[s],[p]and[t]wereusedinsteadtoavoid
86
phonotacticconstraintsinGerman:word-finally,[g]inGermanisproducedas[k]dueto
word-finaldevoicing,and[s]isproducedas[z]inword-initialposition.[æ]isalsousually
used,butwasreplacedwith[ɛ]because[æ]isnotavowelofGerman.[m]and[n]servedas
theexperimentallyrestrictedconsonantsfollowingpreviousexperiments(Delletal.,2000;
Warker,2013;WarkerandDell,2006).Anothersetofconsonants([k],[p],[t],[f])was
unrestricted,appearingfreelyinanyposition.[ŋ]and[h]servedasthelanguage-wide
restrictedconsonants,with[ŋ]illegalinonsetand[h]illegalincoda.
Eachmodeltalkerrecordedasetof96uniquetonguetwisterstobeusedinthe
DifferentandSharedconditions(thesetwaskeptconstantacrossmodeltalkers).These
twisterswererandomlyconstructed,withtheexceptionofphonotacticconstraints:they
obeyedlanguage-wideconstraints,andhalfofthesetwistershad[m]inonsetand[n]in
coda,whiletheotherhalfhadthereversepattern.Thephonotacticpatternassignedtoeach
modeltalkerwascounter-balancedacrossparticipant.IntheDifferentandShared
conditions,participantswereexposedtoalternatingmodeltalkers(andtherefore
alternatingphonotacticconstraints)oneverytrial.The[ɛ]vowelwasusedineachofthese
twisters.Forexample,aparticipantinoneoftheSharedorDifferentconditionsmightbe
exposedtothefollowingtwotrials:
Trial1(Shared/Differentconditions)
TalkerA:fengmethepken
Trial2(Shared/Differentconditions):
87
TalkerB:nengfekhempet
Notethatthepositionof[m]and[n]isdependentonthetalker,whichalternateseachtrial,
whilethevowelisconsistentbetweentrials.
Withtheexceptionofthealternation,thetwisterorderwasrandomized.Ineach
session,participantswereexposedtothesetof96twisters—halffromeach
talker/phonotacticconstraint—thenexposedtothesamesetagaininadifferentrandom
order,foratotalof192trials.Therewasnoindicationforparticipantsthattheywere
exposedtoasmallersettwice,ratherthanonelargeset.IntheVowelcondition,
participantsheardasinglemodeltalker,witheachtrialalternatingbetweenthe[ɪ]and[ɛ]
vowels,withthephonotacticconstraintalternatingbasedonthevowel.Forexample,a
participantintheVowelconditionmightbeexposedtothefollowingtwotrials:
Trial1(Vowelcondition)
TalkerA:tikminpifhing
Trial2(Vowelcondition):
TalkerA:kepnetfenghem
Notethatthepositionof[m]and[n]aredependentontheidentityofthevowel,which
alternateseachtrial,whilethetalkerisconsistentthroughouttheexperiment.
Eachmodeltalkerrecordedanadditional48twisterswiththe[ɪ]vowelforthe
Vowelcondition,withtheother48twistersre-usedfromtheotherconditions.
88
3.3.3. Procedure
Participantswererecordedinasound-proofboothusingahead-mounted
microphone,wheretheyreadthetwistersfromacomputermonitorandheardtwisters
frommodeltalkersoverspeakers.Inaddition,participantsheardmetronomebeatsfrom
anearbudheadphoneplacedintheleftear.
Beforetheexperimentaltrialsbeganeachsession,participantscompleted4practice
trials,noneofwhichcontainedtheexperimentallyconstrainedconsonants.Practicetrials
includedthemodeltalkersthatwouldappearintheexperimentaltrials.Ineach
experimentaltrial,asequenceoffoursyllableswaspresentedtoparticipants(e.g.,femheng
ketnep).Intheslowrepetitionphaseofeachtrial,participantswereexposedtothe
sequencebeingreadbythemodeltalkerataslowpace(1syllable/second).Afterthe
modeltalkeraudiostoppedplaying,theorthographicrepresentationofthetwister
immediatelyappearedonthescreen,andparticipantsrepeatedthetwisterintimewiththe
modeltalker.Eightsecondsafterinitiatingthetrial(4secondsforthemodeltalker’s
productions,and4secondsfortheparticipantsrepetition),participantswerefreeto
advancetheexperimenttothefastrepetitionphase.Inthefastrepetitionphase,
participantsheardsixteenbeatsofametronomeatthespeedof2.5syllables/second.
Participantswereinstructedtowaitduringthefirst4beats,thenrepeatthefour-syllable
sequencethreetimesoverthefinal12beats.Afterallmetronomebeatswerefinished
playing,participantswerefreetoadvancetheexperimenttothenexttrial.
Afterthefinalsession,participantswereaskedaseriesofpost-experiment
89
questionsbytheexperimenter:first,didyounoticeanythingnoteworthyaboutthe
speakers?Ifparticipantsdidnotbringupthenon-nativestatusofspeakersontheirown
(intheconditionsinwhichtheywereexposedtonon-nativespeakers),theywereaskeda
secondquestion:didyouthinkthespeakersyouheardwerenativespeakersofEnglishor
non-nativespeakers?Finally,ifparticipantsrespondedthattheyheardoneormorenon-
nativespeakers,theywereaskedtoguessthespeakers’nativelanguagebackgrounds.
3.3.4. Analysis
Adaptationwasstatisticallyverifiedusinglogisticmixed-effectsregressions.The
dependentmeasureinallmodelswasmaintenanceofsyllableposition.Thefirsttypeof
analysisfittedindividualmodelstothedatafromeachcondition.Fixedeffectsincludeda
contrast-codedfixedeffectofconsonanttype(i.e.,experimentallyconstrainedvs.
unrestricted)andacontrast-codedfixedeffectofexperimentalsession,inwhichsession1
wascontrastedwithsessions2and3.Aninteractiontermbetweensessionandconsonant
typewasalsobeincluded.Thisinteractiontermwouldindicateadaptationhadtakenplace,
withanincreasedmaintenanceofsyllablepositionforconstrainedconsonantsinsessions
2and3oversession1.Randomeffectswillincluderandominterceptsfortargetsyllable;
withonly4participantspercondition,therewasnotasufficientnumbertoincluderandom
effectsforparticipant.
Inacomparisonacrossconditions,amodelwasfittedtothedatafromtheDifferent
andSharedconditions(excludingtheVowelcondition).Thismodelincludedcontrast-coded
90
fixedeffectsconsonanttype,session,andtheirinteraction,andrandominterceptsfor
targetsyllable.Inaddition,itincludedacontrast-codedfixedeffectofcondition,comparing
theSharedconditionswiththeDifferentcondition.Ifspeakersmaketop-downinferences
aboutlanguagebackgroundduringproduction,wewouldexpectasignificantinteraction
suchthattheeffectofconsonanttypeisstrongerintheDifferentconditionthaninthe
Sharedconditions.
Asecondcontrast-codedterm,comparingtheNativeSharedconditiontotheNon-Native
conditionandtheDifferentcondition(i.e.,conditionswithonlynativemodeltalksvs.those
withsomeorallnon-nativemodeltalkers),wasplannedbutultimatelynotincluded.
Finally,alternativeversionsofthewithin-conditionmodelsandbetween-condition
modelswerefittedtothedatajustfromsessions2and3.Thestructureofthesemodels
wereidenticaltopreviousmodels,withtheexceptionofsession,whichwasnotincludedas
afixedeffect.Modelswithoutsession1wereunplannedanalyses;theirpurposewasto
simplifymodelstructureandfocusonsessions2and3,wherewehadthestrongestprior
beliefthatadaptationwouldoccur.
3.4. Results
Recordingsweretranscribedforerrors.Errorsmadeupofconsonantsnotincluded
intheexperiment(N=64)werenotedbutexcludedfromtheanalysis;individualvowel
errorsweregenerallynotnoted,butifparticipantsmadesystematicvowelerrorsthiswas
noted.Errorswerecodedformaintainingsyllableposition.Forexample,ifthetarget
twisterwasfektephenmengandtheparticipantproducedtektephenmeng,the[t]errorin
91
tekwascodedasmaintainingitsposition;iftheparticipantproducedfettephenmeng,the
[t]errorinfetwascodedaschangingitsposition.Everytwisterwastreatedasifithad
exactly4targetsyllables.Ifparticipantsproducedmoreorlessthan4syllables,participant
productionswerealignedwiththe4twistertargetsinthealignmentthatresultedinthe
fewesterrors.Errorswerealsocodedforraterconfidenceaseitherhighconfidence(i.e.,
verylikelyorcertainerror)orlowconfidence(i.e.,possible/moderatelylikelyerror).
Acrossallconditions,therewere48totalsessions(16participants,3sessionseach),
resultinginatotalof9,216trials(2304percondition)or110,592totalsyllables(27,648
percondition).Errorswerecodedby2coders;ofthe48total,onecodercompleted43
sessions,whileasecondcodercompleted6sessions.Onesession(2,304syllables)was
codedforreliabilitybybothcoders.Therewasanoverallagreementrateonthe
presence/absenceoferrorsfor99.3%ofallsyllables.Lookingatagreementononlythose
errorsidentifiedbytheprincipalcoder(N=27),therewas74.1%conditionalized
agreement,anumberinlinewithpreviousexperiments.
Atotalof3864errorswerediscovered,or3.5%ofallsyllables,anerrorratewithin
therangeofpreviousexperiments(althoughonthelowend;seeAppendixF).Ofthese,82
(2.1%)werecodedaslowconfidencebythetranscribers.Arelativelywiderangeoferror
rateswerefoundfordifferentconditions(seeTable3.2),butthatmaysimplyreflect
variationacrossparticipants(seeFigure3.1).
92
Table3.2.Numberoferrorsanderrorratesbycondition.
Condition nErrors ErrorRateDifferent 928 3.4%NativeShared 1201 4.3%Non-NativeShared 1106 4.0%Vowel 629 2.3%
Figure3.1.Overallerrorratesbyparticipant.Colorsreflectexperimentalcondition.
Therewere807errorsonconsonantssubjecttolanguage-widephonotactic
constraints(i.e.,errorsresultingin[ŋ]and[h]).Ofthese,99.8%followedEnglish
phonotacticconstraintsandmaintainedtheirsyllablepositions.Therewere2,124errors
forunconstrainedconsonants([p],[t],[k],and[f]),whichmaintainedtheirsyllableposition
74.5%oftime,inlinewithpreviousexperiments.Ofthe933errorsforexperimentally
constrainedconsonants([m]and[n]),however,only50.2%oferrorsmaintainedtheir
93
syllableposition.Thiswasasurprisingresult—evenifnolearningoccurredatall,
experimentallyconstrainedconsonantsshouldhave,allelsebeingequal,maintainedtheir
syllablepositionatasimilarratetounconstrainedconsonants.Closerinspectionofillegal
constrainederrorsrevealedthatalargeproportionconsistedofswapsbetween[m]and
[n]:for40.5%ofall[n]errors,thetargetwas[m],andfor32.7%ofall[m]errors,thetarget
was[n](seeTable3.3fortarget-errormatrix).Overall,swapsbetweennasalconsonants
occurredatadisproportionatelyhighrate:tabulatingtheproportionoferrorsthatwere
intendedforaspecifictargetconsonantforall64possibleerror/targetcombinations
revealedthat4ofthetop5error/targetcombinationswerenasalswaps.Thelocusforsuch
swapsmaybethatnasalpairsweremorephoneticallysimilaritythanotherconsonant
pairs;previousevidencefromtonguetwisterexperimentssuggestsphoneticsimilarity
(Wilshire,1999)andoverlappingphonologicalfeatures(Goldrick,2004)affectspeech
errors.It’sunclear,however,why[m]and[n]weremoreaffectedbyphoneticsimilarity
thanothersegments,asanumberofpreviousexperimentshaveincludedthesesegments
(seefollowingsectionforfurtherdiscussion).
94
Table3.3.Target/errormatrixforallconsonantsandconditions.Targetsarecolumnsanderrorsarerows.TheNoTargetcolumnreferstoerrorsthatweremadeonextrasyllables(i.e.,whenaparticipantproducedmorethan4syllablesinasingletwister).Gradientcolor-codingreflectsthenumberoferrorsforagiventarget/errorcombination.
ftargets
htargets
ktargets
mtargets
ntargets
ŋtargets
ptargets
ttargets
NoTarget
ferrors - 69 85 44 19 12 133 33 29herrors 55 - 129 32 54 0 52 51 32kerrors 103 158 - 20 11 35 150 126 23merrors 24 18 22 - 119 99 50 12 20nerrors 16 39 27 231 - 188 13 34 21ŋerrors 10 0 22 106 231 - 8 12 13perrors 105 75 174 62 17 5 - 96 31terrors 44 57 167 9 35 3 163 - 31
Such[m]-[n]swapshadalargenegativeeffectonthemaintenanceofsyllable
positionbecause[m]and[n]targetsneveroccurredinthesamesyllablepositioninagiven
twister(i.e.,[m]-[n]swapswerealwaysillegal).Assuch,inadditiontoanalysesconducted
withthefulldataset,post-hocanalyseswerealsoconductedwith[m]-[n]swapsexcluded
fromthedatasettocontrolforthisphoneticsimilarityeffect(Goldrick,2004,followeda
similarprocedure).Thegoaloftheseanalyseswastocompareeffectsbetweenconditions,
oncephoneticsimilaritywascontrolled.Notethatwhileothernasalswapsinvolving[ŋ]
alsooccurredatadisproportionatelyhighrate,suchswapswouldnothavethesameeffect
onthesyllablepositioneffectbecause[ŋ]targetsalwaysoccurredincodaposition.Unlike
[m]-[n]swaps,nasalswapsinvolving[ŋ]andanothernasaldidnot,bydefinition,change
syllablepositions.Assuch,swapsincluding[ŋ]wereatnopointexcludedfromthedataset.
Inaddition,giventheissueswithperceptualsimilarity,low-confidenceerrorswere
excluded.
95
Thepercentageoferrorsthatmaintainedsyllableposition,splitbyexperimental
sessionandcondition,isshowninFigure3.2.Inwithin-conditionmodelsthatincludedall
sessions,thecriticalinteractionwasbetweentheconstrainttermandthesessionterm.No
suchinteractionsweresignificantforanyconditionexcepttheNativeSharedcondition,
whichwentintheoppositedirectionofadaptation(i.e.,thesyllable-maintenanceeffected
decreasedforexperimentallyconstrainedconsonantsinsessions2and3;seeAppendixG
forfullmodelresults).Contratoourpredictions,adaptationdidnotincreaseinlater
sessionswithineithertheVowelorDifferentconditions(seebelowfordiscussionofnull
resultforthebaselineVowelcondition).
96
Figure3.2.Percentageoferrorsthatmaintaintheirsyllablepositionforconstrainedvs.unconstrainedconsonants,brokendownbysessionandcondition.Alldatavisualized,including[m]-[n]swaps.Errorbarsreflect95%confidenceintervaloverparticipant;however,notethatparticipantscontributeddifferentnumbersoferrorstoeachbar.
Inthebetween-conditionmodelthatincludedallsessionsandcomparedadaptation
intheDifferentconditiontotheSharedconditions,thecriticalinteractionbetween
97
constraint,session,andconditionwasnotsignificant(seeTable3.4)8.Thissuggeststhat,
againstourprediction,adaptationtotheexperimentalconstraintdidnotincreasemorefor
theDifferentconditioninlatersessionsthanfortheSharedconditions.Inthebetween-
conditionmodelthatonlyincludedsessions2and3,however,thecriticalinteraction
betweenconstraintandconditionwassignificant,suggestingthatforthelatersessions
only,therewasalargeradaptationeffectintheDifferentconditionthanintheShared
conditions(seeTable3.5).Thisprovidesevidenceforastrongereffectofadaptationinthe
latersessionsoftheDifferentconditionthanintheSharedconditions,althoughitshouldbe
notedthiswasanunplannedanalysis.
Table3.4.Betweenconditioncomparison;allsessions;[m]-[n]switchesincluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.24 0.11
Constraint 0.3* 0.14 4.72 0.03Session 0.03 0.13 0.05 0.83Condition -0.18 0.19 0.85 0.36Constraint:Session -0.12 0.26 0.23 0.63Constraint:Condition 0.35 0.38 0.87 0.35Session:Condition 0.46 0.38 1.44 0.23Constraint:Session:Condition 0.9 0.77 1.36 0.24
RandomEffects
VarianceTargetSyllable
0.31
8Itshouldbenotedthatthiscriticalinteractionwasmarginallysignificant(p=0.07)whenlow-confidenceerrorswereexcluded,illustratingthesensitivityoftheeffecttovariousanalysischoices—seebelowforfurtherdiscussion.
98
Table3.5.Betweenconditioncomparison;session1excluded;[m]-[n]switchesincluded
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 0.5 0.12
Constraint -1.12*** 0.15 58.39 <0.001Condition 0.05 0.19 0.06 0.8Constraint:Condition 0.81* 0.39 4.32 0.04
RandomEffects
VarianceTargetSyllable
0.42
Inthepost-hocanalysis,all[m]-[n]switcheswereremovedfromthedataset(see
Figure3.3).Thishasalargepositiveeffectonthesyllablemaintenanceeffectfor
experimentallyconstrainedconsonants(raisingitfrom50.2%to80.3%),bringingitmuch
closertothesyllablemaintenanceeffectforunconstrainedconsonants(74.5%).Inthe
within-conditionmodelsthatincludedallsessions,butexcluded[m]-[n]swaps,thecritical
constraint:sessioninteractionwasnotsignificantwithinanycondition.Thissuggeststhat
evenwhenremoving[m]-[n]swaps,thereisnoevidencethatsyllablemaintenance
increasesaftersession1moreforconstrainedconsonantsthanforunconstrained
consonants(seeAppendixGformodelresults).
99
Figure3.3.Percentageoferrorsthatmaintaintheirsyllablepositionforconstrainedvs.unconstrainedconsonants,brokendownbysessionandcondition.[m]-[n]swapswereexcluded.Errorbarsreflect95%confidenceintervaloverparticipant;however,notethatparticipantscontributeddifferentnumbersoferrorstoeachbar.IntheVowelcondition,oneparticipantmadethemajorityoftheerrorsbuthadamuchlowermeanthanotherparticipants,resultinginaCIthatdoesnotoverlapwiththemeanforSession1,unconstrainederrors.
Inaseparatewithin-conditionanalysisthatexcludessession1,thecriticalmain
effectofconstraintwassignificantintheDifferentcondition(seeTable3.6),butnotinany
othercondition(seeAppendixG).Thissuggeststhatonce[m]-[n]swapswereremoved,
syllablepositionwasmaintainedatahigherrateforconstrainedconsonantsvs.
100
unconstrainedconsonantswithsessions2and3.Importantly,thiseffectdoesnotemerge
foranyothercondition,providingsomelimitedevidenceforgreateradaptationinthe
Differentcondition.
Table3.6.Differentcondition;session1excluded;[m]-[n]switchesexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.23 0.2
Constraint 0.73* 0.35 4.76 0.03
RandomEffects
VarianceTargetSyllable
0.81
Finally,abetween-conditionanalysiswasperformedonthedataexcluding[m]-[n]
swaps.Inamodelthatincludedallsessions,thecriticalthree-wayinteractionbetween
condition,constraint,andsessionwasnotsignificant(seeTable3.7).Inamodelthatonly
includedsessions2and3,thecriticaltwo-wayinteractionwasalsonotsignificant(see
Table3.8),differingfromtheeffectfoundwhen[m]and[n]wereincludedintheanalysis.
101
Table3.7.Betweenconditioncomparison;allsessions;[m]-[n]switchesexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.23 0.11
Constraint 0.32* 0.14 5.48 0.02Session 0.01 0.13 0.01 0.92Condition -0.19 0.19 0.99 0.32Constraint:Session -0.08 0.26 0.09 0.76Constraint:Condition 0.37 0.38 0.95 0.33Session:Condition 0.49 0.38 1.61 0.2Constraint:Session:Condition 0.81 0.76 1.12 0.29
RandomEffects
VarianceTargetSyllable
0.31
Table3.8.Betweenconditioncomparison;session1excluded;[m]-[n]switchesexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.26 0.13
Constraint 0.23 0.19 1.55 0.21Condition 0.05 0.26 0.04 0.85Constraint:Condition 0.88 0.52 2.93 0.09
RandomEffects
VarianceTargetSyllable
0.29
Thepost-testquestionsrevealedthatparticipantsgenerallyhadlittleexplicit
knowledgeaboutthedesignoftheexperiment.Noparticipantsindicatedtheywereaware
ofthephonotacticconstraintsbasedontalker/vowel.Only2participants,ofthe10who
wereexposedtoatleastonenon-nativetalker,wereabletodetectthattheyhadhearda
non-nativetalker.Ofthese,oneheardthenon-nativetalkerasbeinganativespeakerofan
EastAsianlanguage,whiletheotherheardthenon-nativetalkerasbeinganativespeaker
ofVietnamese.It’sunclearwhetherexplicitdetectionofmodeltalker’snon-nativelanguage
102
backgroundisnecessarytosuccessfullyadaptintheDifferentcondition.(Notethat
previoustonguetwisterexperimentshaveshownthatexplicitknowledgeofthe
experimentaldesignhavemadelittledifferentinparticipantbehavior;e.g.,Dell,etal.
2000.)
Tosummarize,wefoundinconsistentevidenceforourinitialpredictions.Two
significantresults,bothfollowingfrompost-hocanalyses,supportingthehypothesiswere
found:first,thatwhen[m]-[n]swapswereincluded,participantsshowedahighersyllable
maintenanceeffectinlatersessionsoftheDifferentconditionthaninthelatersessionsof
theSharedconditions.Andsecond,thatwhen[m]-[n]swapswerenotincluded,
constrainedconsonantsmaintainedtheirpositionmorefrequentlythanunconstrained
consonantswithinthelatersessionsoftheDifferentcondition(andnotinanyother
condition).Overall,however,theseeffectswerequitebrittle,andchangeddependingon
whichsessionswereincludedintheanalysisandwhetherornot[m]-[n]swapswere
included.Moreover,wesurprisinglydidnotfindsignificantresultofadaptationinthe
baselineVowelcondition,whichmakesitdifficulttocontextualizetheresultsinother
contiditions.
3.4.1. Discussion
Inthecurrentstudy,participantswereexposedtonovel,non-nativephonotactic
constraintscontingentonmodeltalkerinatonguetwisterparadigm,withthelanguage
backgroundofmodeltalkersmodulatedacrossconditions.Theresultswereultimately
inconclusive,withsomeweakevidencepointingtowardsparticipantsacquiringthe
103
constraintswhenmodeltalkersdifferedintheirlanguagebackground,butnotwhenthey
sharedalanguagebackground.Inaddition,weunexpectedlydidnotreplicatethewell-
establishedadaptationeffectsfromprevioustonguetwisterstudiesintheVowelcondition,
whichmakescontextualizingresultsfromotherconditionsdifficult.Finally,theoverall
syllablemaintenanceeffectforconstrainedconsonantswassurprisinglylowduetoalarge
numberof[m]-[n]swaps,whichfurtherobscuredanycomparisonsbetweenconditions.
Whatisbehindthehighnumberof[m]-[n]swaps?Giventhatmultipleprevious
studieshaveusedtheseconsonantswithoutsuchcomplications,it’slikelythatthe
differenceinthemethodologyofthecurrentstudy—inwhichparticipantslistentothe
tonguetwisterbeforeproducingit—contributedtotheabnormallyhighnumberofswaps.
Onepossibility,giventhatthistaskinvolvedperception,isthehighperceptual
confusabilityof[m]and[n]:listenersfrequentlymisidentify[m]as[n],andvice-versa,in
speechinnoise(e.g.,MillerandNicely,1955;PhatakandAllen,2007).It’spossiblethatthe
similarityof[m]and[n]inperceptionresultedindown-streamerrorsinproduction.While
otherconsonantsincludedintheexperimentarealsoperceptuallyconfusable(e.g.,[p]and
[k];seeTable3.3),swapsbetweenunconstrainedconsonantswerenot,bydefinition,
alwaysillegal,andthuswouldnothaveaffectedsyllablemaintenanceratesinthesame
way.
Anothercontributingfactormayhavebeenthecoder’sperceptionofparticipants’
[m]and[n]productions.Ifthecoderhadanatypical[m]-[n]categoryboundary,inwhich
oneperceptualcategorywaswideandtheotherwasfairlynarrow,itcouldhaveresultedin
frequentmiscategorizations.Specifically,wewouldexpectthecodertoconsistentlychoose
104
onenasalovertheother.Indeed,therewere231total[n]errorsfor[m]targets,andonly
119[m]errorsfor[n]targets(Miller&Nicely,1955,findasimilarasymmetryin
perception).Thiswasfairlyconsistentacrosscoders,however,suggestingit’slikelynot
relatedtoonecoder’sidiosyncraticperceptualcategoryboundary.
Smalleetal.,(2017)—theonlyotherstudytouseasimilarparadigm—didnotfind
depressedsyllablemaintenanceeffectsforconstrainedsyllables.Instead,theauthorsfound
asurprisinglyhighsyllable-positioneffectforunconstrainedsyllables(87.4%,roughly
12%higherthanpreviousstudies),whichwasnotreplicatedinthisstudy.Therewerea
numberofimportantdifferencesbetweenthecurrentstudyandSmalleetal.,(2017)that
mighthaveresultedinthedivergentfindings(differentsetofconsonantsandvowels;
Dutchparticipants/twisters,ratherthanEnglish;4sessions;etc.)
AnotherperplexingfindingwasthelackofaneffectintheVowelcondition.When
[m]-[n]swapsareremoved,thesyllablepositioneffectdoesreach100%forconstrained
consonantsinthethirdsession,whichisinlinewithpreviousresults.Thisisoveronly13
errors,however,whichisfartoosmallasampletodrawdefinitiveconclusionsfrom.This
smallsamplesizemayhaveresultedinthelackofaneffect—evenwith[m]-[n]swaps
included,the629errorswouldbethesecondsmallestnumberreportedforanyprevious
tonguetwisterexperimentusing2nd-orderconstraintsconditionedonvowels(see
AppendixF).ThiswasinlargepartduetotheverylowerrorrateintheVowelcondition
(2.3%)—lowerthananypreviousexperimentwith2nd-orderconstraintsexceptSmalleet
al.,(2017).Giventhattheerrorratefortheexperimentasawholewasinlinewith
previouswork,thelowerrorrateintheVowelconditionwasmostlikelycausedbythe
105
smallnumberofparticipants,andlargevariancebetweenparticipanterrorrate:thetwo
participantswiththelowesterrorrates,bothlessthan1%,wereintheVowelcondition.A
relatedproblemwasthatanotherparticipantintheVowelconditionhadthesecondhighest
errorrate.Whilethisincreasedthesizeofthesample,62%oftheerrorsfortheVowel
conditioncamefromasingleparticipant,whichisclearlynotarepresentativesample.
AsecondpossibleissuewiththeVowelconditionwasthephoneticsimilarityofthe
vowels—[ɪ]and[ɛ]aremuchmoresimilarthan[ɪ]and[æ],thevowelsusuallychosenin
previoustonguetwisterparadigms.Inthepost-experimentquestionnaire,oneparticipant
mentionedtheycouldnotconsistentlytellthemodeltalker’svowelproductionsapart,
whiletwootherssometimesheardthemodeltalkers’[ɛ]vowelas[ɪ].Giventhatthe
phonologicalfeaturesoftheconsonantsinvolvedina1st-orderconstraintaffectserrorsin
tonguetwisters(Goldrick,2004),itmayalsobethecasethatthesimilarityofthe
conditioningvowelsina2nd-orderconstraintcanweakenthesyllable-positioneffect.
Infuturefollow-upexperimentsusingasimilardesign,anumberofstepscanbe
takentoavoidsomeoftheissueshighlightedabove.Mostobviously,alargernumberof
participantsisrequiredforamorerepresentativesample.Thatsaid,therearesome
logisticalobstaclestoincreasingthenumberofparticipants:becausetherecordingsneed
tobecodedforerrorsbyhand,itisahighlytimeandmoney-intensiveparadigm.One
strategytomitigatethiscostistoshortenthelengthoftheexperiment.Asinprevious
tonguetwisterexperiments,thenumberoferrorsdecreasesoverthecourseofthe
experiment,bothwithineachsessionandacrosssessions—roughlyhalfofallerrors
occurredinthefirstsessionalone—resultingindiminishingreturns.Shorteningthelength
106
oftheexperimentto96trialspersessionandreducingthenumberofsessionsto2,while
increasingthenumberofparticipants,willresultinmoreefficientspeecherrorcoding,due
totheincreaseddensityoferrors.Second,afollow-upstudyshoulduselessperceptually
confusableexperimentallyconstrainedconsonants.Whilemostpreviousstudieshaveused
phoneticallysimilarconsonants(e.g.,[k]-[g],[m]-[n],[f]-[s]),thereisnoapriorireasonfor
this.Movingawayfromsuchpairsshouldmitigatethepossibleinteractionbetweenthe
auditorypresentationofthetwistersandthephoneticsimilarityofconstrainedsegments,
resultinginfewerillegalswitchesbetweenexperimentallyconstrainedconsonants.
It’spossibleadifferentdesignmaybebettersuitedtoinvestigatethequestionat
hand.Onealternativeisamodificationoftheconstraint-switchingdesignemployedin
Andersonetal.(2019).Inthefirstexperimentalblock,participantswereexposedtoa1st-
orderconstraintinatonguetwisterexperiment.Thisconstraintwasthenreversedinthe
secondblock(e.g.,if[f]wasconstrainedtoonsetinBlock1,itwasconstrainedtocodain
Block2),thenreversedagaininthethirdblock.Andersonandcolleaguesfoundslower
adaptationforthereversaloftheconstraint,suggestinganincrementalphonotactic
learningmechanism.Undertheframeworkpresentedinthisstudy,however,participants
werelearningandthenunlearningthesamelaboratory“mini-grammar”inallblocks.Ina
modifiedversionofthestudy,twisterscouldbepresentedauditorily,withthemodel
talkersswitchingtheirlanguagebackgroundsbetweenblocks(seeWeissetal.,2009,fora
similardesigninastatisticallearningparadigm).Thismayserveasacuetolearnersthat
theyareinfactbeingexposedtodifferentlanguagesineachblock,spurringfaster
adaptation,encouragingthemtoseparatetheevidencefortheconflictingconstraintsinto
107
different“mini-grammars”.Inaddition,thisparadigmonlyrequiresasinglesession,
allowingfortherecruitmentofagreaternumberofparticipants.
3.4.2. Production-perceptiondynamicsandphonotacticadaptation
Recentresearchexploringtherelationshipbetweenspeechproductionandspeech
perceptionhasplacedalargeemphasisontheroleofprediction(e.g.,DellandChang,2014;
PickeringandGarrod,2007;PickeringandGarrod,2013).PickeringandGarrod(2007)
proposethatwhentheperceptualsystemmakespredictions—aubiquitousprocessin
comprehensionthatoccursatmultiplelevelsofrepresentation—itrecruitstheproduction
systemtoconstructaforwardmodelthatanticipatesupcominglinguisticinput.The
discrepancies,orerror,betweentheforwardmodel’spredictionsandtheobservedinput
drivefutureadaptation.Underthisframework,phoneticimitationisaresultofcovert
imitationduringperception,duringwhichspeakersrecruittheproductionsystemto
predicttheirinterlocutor’sspeech.Thisinturnbringsthespeaker’sproductionsystem
moreinlinewiththeirinterlocutor’s,resultinginphoneticimitation.
KittredgeandDell(2016)investigatedtransferofphonotacticconstraintslearnedin
perceptiontothoselearnedinproduction.Participantsalternatedproducingtongue
twistersandhearingstringsofsyllablesthatconflictedintheirphonotacticconstraints.
Critically,constraintsparticipantswereexposedtoinperceptiononlyinterferedwith
constraintsinproductionwhentheperceptiontaskinvolvedimitation,eitherbywayof
silentproductionorerrormonitoringwithoutorthographicsupport(seealsoWarkeretal.,
108
2009).
Inthecurrentstudy,participantsshadowedthemodeltalker’sspeech.Under
PickeringandGarrod’sframework,ifthisshadowinginvolvescovertimitation,and
thereforeprediction,itshouldhaverecruitedtheproductionsystem.Whiletheresultsof
thecurrentexperimentaresomewhatinconclusive,it’spossiblethislinkbetween
perceptionandproductionisacriticalpathwaytopassinformationaboutcausal
inferences—inthiscasetherelationshipbetweenphonotacticvariationandthemodel
speakers’backgrounds—fromtheperceptionsystemtotheproductionsystem.It’spossible
thatwhiletheproductionsysteminisolationisrelativelyinsensitivetothecausalstructure
underlyingphonotacticvariation(e.g.,Andersonetal.,2019),whenitisco-activewiththe
perceptionsystem,itgainsaccesstosuchcausalinferences.Theinconclusivecurrent
resultsdonotprovideevidenceforthisaccount,however,andfollow-upstudiesarecritical
tofurtherexplorethispossibility.
3.5. Conclusion
Inthecurrentstudy,weinvestigatedtheroleofcausalinferenceinphonotactic
adaptationinatonguetwisterparadigm.Wepositedthatspeakerswouldusetheirprior
experiencewithphonotacticvariation—thatitvariestoalargedegreebetweenlanguages,
andverylittlebetweenspeakersofthesamelanguagevariety—toguidetheiradaptationto
novelconstraints.Weexposedparticipantsto2nd-orderphonotacticconstraints
conditionedonmodeltalker,whilemodulatingthelanguagebackgroundsofthemodel
109
talkers.Wepredictedthatparticipantswouldadapttoagreaterdegreewhentalkers
differedintheirlanguagebackgrounds,astheywouldconstructaseparatephonotactic
grammarforeachtalker(i.e.,theywoulddetecttheywerelearningtwoseparate
“laboratorylanguages”).Theresultswereultimatelyinconclusive,althoughsomeevidence
wasfoundinthepredicteddirection.Totheextenttheseeffectsarereliable,itsuggests
thatparticipantsmayusethelanguagebackgroundofthemodeltalkersasacuetoa
changeinphonotacticcontextbetweenmodeltalkers.Iftheseeffectswereconfirmedin
futurestudies,itwouldsuggestthatspeechproductionismoresensitivetosuchhigh-level
causalinferencesthanrecentevidencehassuggested,especiallyinferencesinvolvingthe
languagebackgroundsoftalkers.
110
4. Conclusion
Inthisdissertation,weexploredthewaysinwhichthestructureofphonotactic
variationthatspeakersexperiencecaninduceadaptationtonovelconstraintsinsome
contexts,anddampenitinothers.Westartfromtheassumptionthatadaptation,across
domains,ismotivatedbypriorexperience.Thisisacomputationallevelaccountof
phonotacticadaptation(Marr,1982):whilepreviousworkhasfocusedonthemechanisms
involvedinphonotacticadaptation(e.g.,Warker&Dell,2006),thisdissertationexamines
whyweadaptinthefirstplace,andunderwhatcircumstances.Thislineofinquiryhas
previouslybeeninvestigatedinthedomainofphonetics(Kraljic,Samuel,&Brennan,2008;
Liu&Jaeger,2018),inwhichlistenersadapttovariationwhentheyhaveevidenceitis
systematicandrelevantforagiventask(i.e.,differencesbetweenindividualtalkersare
relevantforrecognizingspeechsounds)butdonotadaptwhentheyhaveevidencethe
variationisincidentalforagiventask(i.e.,disruptionsfromapeninthemouthofatalker
arenotrelevantforrecognizingspeechsoundsinothercontexts).Thispatternof
adaptationsuggeststhatlistenersmustproperlyattributevariationtoitsunderlying
sourceforthegiventaskwhenadapting.Inthecaseofphonetics,listenersmakecausal
inferencesaboutthesourcesofvariationbasedonthehighdegreeofvariationendemicto
thephoneticsofindividualtalkers,acontributingfactortotheclassic“lackofinvariance”
problem.
111
Thestructureofphonotacticvariationismarkedlydifferentfromthatofphonetic
variation:individualtalkersarelikelytohavesimilarphonotacticgrammars,whiletalkers
ofdifferentlanguagesarelikelytohavedistinctgrammars.Assuch,wearguedthat
phonotacticadaptationwouldbehavedifferentlyfromphoneticadaptation.Wepredicted
speakerswouldadapttodistinctphonotacticgrammarsfordifferenttalkerstoagreater
degreewhenthosetalkersdifferedintheirlanguagebackgroundthanwhentheyshareda
languagebackground.Ineachexperimentreportedinthedissertation,thispredictionwas
testedbyexposingparticipantstotwotalkersexhibitingdifferentphonotacticconstraints
(i.e.,2nd-orderconstraintsconditionedontalkeridentity),whilemodulatingthelanguage
backgroundofeachtalker.Weexploredadaptationinbothperception(Study1)and
production(Study2);herewesummarizetheresultsofeachstudy,considertheir
implicationsfortheoriesofadaptation,andpositfuturedirectionsforthislineofresearch.
4.1. Study1
InStudy1,wetestedadaptationinarecognitionmemoryparadigm,exposing
English-speakinglistenerstononsensesyllablesthatreflecteddifferenttalker-dependent
phonotacticconstraints(e.g.,forTalkerA,stopsareconstrainedtoonsetandfricativesto
coda;forTalkerB,viceversa).Crucially,talkerseithersharedalanguagebackground(two
FrenchortwoEnglishtalkers),ordifferedintheirlanguagebackgrounds(oneEnglish,one
Frenchtalker).InExperiment1A,wealsomodulatedthestrengthofthecuetothenon-
nativespeakers’languagebackgrounds,withmoreorlessnativeEnglish-likevowels([u]
vs.[y]).
112
Experiment1ArevealedalowdegreeofadaptationfortheNativeSharedcondition,
andamoderatedegreeofadaptationforeveryothercondition.Thisresultlargely
supportedourhypothesis,withagreaterdegreeoflearningfortheDifferentconditions
thantheNativeSharedconditions.Therewere,however,aspectsoftheresultsthatwere
unexpected:first,adaptationoccurredineachcondition,eventhoseinwhichbothtalkers
sharedalanguagebackground.Ouroriginalpredictionwasthatnoadaptationshouldoccur
insuchacontext,asthesharednativebackgroundsofthetwotalkersshouldblock
adaptation.Itappearsthatthestrengthofthebottom-upevidenceoftalker-specific
phonotacticconstraints,however,overcamelisteners’top-downinferencesaboutthe
sourceofthevariation—but,critically,adaptationwasstilllowerthaninotherconditions.
Asecondunexpectedresultwasthatthestrengthofthecuetolanguagebackgrounddid
notappeartoaffectadaptation.Wearguedthattheweakcuewassufficientforlistenersto
determinethenon-nativestatusofthetalker,makingfurthercuesredundant.
Athirdunexpectedresult,andthemosttroublingforthehypothesis,wasthesimilar
degreeofadaptationfortheDifferentconditionsandtheNon-NativeSharedconditions.We
suspectedthissurprisinglyhighdegreeofadaptationfortheNon-NativeSharedcondition
wasaresultofthetwoFrenchtalkerssoundingdissimilar,duetoonetalker’ssomewhat
aberrantproductions.Assuch,inExperiment1Bwerecordedanoveltalker,andre-ran
threeofthefourconditionsof1A(wedidnotincludetheNativeSharedcondition).
Theresultsofthisfollow-upexperimentstrengthenedtheevidenceforour
hypothesis,astheDifferentconditionsshowedahighdegreeofadaptation,andtheNon-
NativeSharedconditionshowedonlyamoderatedegreeofadaptation.Butthisresultalso
113
hadtwosomewhatsurprisingaspects:first,whydidadaptationincreaseintheDifferent
conditions?Itappearsthatparticipantsspecificallydidnotstronglyadapttoaberrant
talker’sphonotacticconstraintsinExperiment1A,regardlessofcondition.Assuch,
replacingthattalkerresultedinanoverallboosttoadaptationintheDifferentconditions.A
secondsurprisewasthatadaptationdidnotdecreaseintheNon-NativeSharedcondition.
Wearguedthatthemoderatedegreeofadaptationinthisconditionwasaresultof
listeners’asymmetricknowledgeofnativevs.non-nativelanguages.Whenlistenersare
exposedtotwotalkersofalanguagetheyarehighlyfamiliarwith,theyarelikelymore
confidentthesetalkerssharealanguagebackgroundthanwhenexposedtotalkersofaless
familiarlanguage.Assuch,theymaybemorelikelytoinferthenon-nativetalkersdon’t
sharealanguagebackground,andadapttoamoderatedegree.
Asawhole,theresultsofExperiment1suggestedthatlistenersadaptedtoagreater
degreewhentalkersdifferedintheirlanguagebackground.Moreover,itsuggested
listenershaveastructuredmodelofphonotactics,assigningdistinctphonotacticgrammars
tonativevs.non-nativelanguages.Itremainedunclear,however,whetherlistenerswere
capableofassigningdistinctphonotacticgrammarstodifferentnon-nativelanguages.In
Experiment2,weaddressedthisquestionbyexposinglistenerstotwonon-nativespeakers
ofdifferentlanguages(HindiandHungarian).Wepositedthatiflistenersmadedistinctions
betweennon-nativelanguages,theywouldtreatanylanguagedifferencesasrelevant
variation,andthereforeadapttoahighdegreeinthiscondition.If,ontheotherhand,
listenerssimplygroupedallnon-nativetalkerstogether,theywouldonlyadapttoa
moderatedegree.Inaddition,weincludedtwoconditionsfromExperiment1:theMixed
114
Differentcondition(i.e.,onenativeandonenon-nativespeaker)andtheNon-NativeShared
condition.
Resultssupportedthewithinnon-nativedistinctionshypothesis:listenersadapted
toahighdegreeofadaptationwhentalkersdifferedintheirlanguagebackgrounds,even
whenbothtalkerswerenon-native.Theseresultssuggestthatthecriticaldistinction
listenersusewhenpositingdifferentphonotacticgrammarsfordifferentspeakersis
languagedifference,ratherthanthemorespecificdistinctionofnativevs.non-native.
Additionally,theNon-NativeSharedconditionreplicatedExperiment1Bwithadifferent
stimulusdesign,languages,talkers,andnumberofitems,showingamoderatedegreeof
adaptation.
Study1shedslightonthemotivationforphonotacticadaptation.Patternsof
adaptationreflectthetypeofphonotacticvariationlistenersexperiencefromvarious
sourcesintheirdailylives,suchasaccentedEnglishspeechorexposuretonon-native
languages.TheseresultsalsosuggestthatphonotacticadaptationisheightenedinL2
contexts(seeWarker,2013),andmaybeakeypartoftheearlystagesofL2acquisition.
Learnersarefacedwithprofoundlydifficultproblems,suchasdiscoveringword
boundaries,atthebeginningofL2acquisition.Rapidphonotacticadaptationinperception,
particularlytosubsetphonotactics,mayserveasafastandefficientadjustmentthatallows
leanerstobettersegmentwords,aspreviousworksuggeststhatphonotacticcuesinthe
inputareusefulforsegmentation(e.g.,Brent&Cartwright,1996)andthatadultsusethese
cuestoguidesegmentation(e.g.,McQueen,1998).AdaptingtoL2subsetphonotacticsmay
allowlearnerstoquicklynarrowthesetofpossibleword-formstheyhavetoconsider
115
duringcomprehension,aidingsegmentationandlexicalaccess.Suchquickadjustmentsto
L2speecharelikelynotavailableforotherphonologicalstructuresthatareheavily
influencedbylow-levelperceptualprocesses.L2supersetphonotactics(e.g.,consonant
clustersnotpresentintheL1),forexample,areverydifficulttolearn(e.g.,Parlato,
Christophe,Hirose,&Dupoux,2010),whichmayreflectinpartperceptualinterference
fromlisteners’L1s(e.g.,Dupoux,etal.,1999).Phonotacticadaptation’sroleinL2
acquisitionisaripeavenueforfutureresearch(seefuturedirectionssectionbelow).
4.2. Study2
InStudy2,weexaminedwhethertheprinciplesofadaptationwetestedin
perception—thatlearnersmakeinferencesaboutthecauseofvariationbasedontheir
priorexperience,andtheseinferencesguideadaptation—alsoholdinspeechproduction,
usingamodifiedtonguetwisterparadigminwhichparticipantsrepeataftermodeltalkers.
Evidencefromprevioustonguetwisterexperimentssuggestedthatspeakersbuiltseparate
phonotacticgrammarsinlaboratorycontexts,andwereabletoseparateevidencefor
constraintsinthelaboratorygrammarfromthoseusedintheirnativelanguage(Warker,
2013).AsinperceptioninStudy1,wearguedthatspeakerswererecruitingtheirL2
acquisitionfacultiesinadaptation.IfadaptationisdrivenbyL2acquisition,thenitfollows
thatlearnersshouldbesensitivetocuesregardingthelanguagebackgroundofthemodel
talkerstheyarerepeatingafter.Thelanguagebackgroundoflistenersmayserveascuefor
learnerstodetectachangeincontext,andseparatetheevidenceforconflictingphonotactic
constraintsintoseparategrammars(e.g.,Weiss,etal.,2009).However,anumberofrecent
116
studieshavesuggestedthatphonotacticadaptationinproductionisresistanttotop-down
inferences(e.g.,Andersonetal.,2019),suggestingthatlearnerswillbeunabletoseparate
evidencefortheconflictingphonotacticconstraintsbasedontalkerlanguagebackground.
MirroringthedesignfromStudy1,participantswereexposedtomodeltalkerswho
sharedanativelanguagebackground(2Englishtalkers),anon-nativelanguage
background(2Germantalkers),ordifferedintheirlanguagebackgrounds(1English,1
Germantalker),witha2nd-orderphonotacticconstraintconditionedontalkeridentity.A
controlconditionwasincludedinwhichthephonotacticconstraintwasconditionedonthe
identityofthevowel,followingprevioustonguetwisterstudies(e.g.,Warker&Dell,2006).
WepredictedahighdegreeoflearningintheDifferentandVowelconditions,andalow-to-
moderatedegreeoflearninginthetwoSharedconditions.
Theresultswereultimatelyinconclusive.Whiletherewassomeevidenceofahigher
degreeoflearningintheDifferentcondition,theeffectwasbrittle,andwasnotconsistent
amongdifferentanalyses.Surprisingly,wefoundnoeffectoflearninginthecontrol
condition,possiblyduetoasmallnumberoferrors,amajorityofwhichcamefromasingle
participant.Therewasalsoanoverwhelmingeffectofphoneticsimilarityforthetarget
consonants([m]and[n])inallconditions,resultinginahighnumberof[m]-[n]swaps.Due
tothedesignoftheexperiment,theseswaps,bydefinition,violatedthephonotactic
constraint,resultinginsurprisinglylowoverallmaintenanceofsyllablepositionforthe
experimentallyconstrainedconsonants.Thismayhavebeenduetothemodifieddesign,in
whichparticipantsheardthetwistersbeforerepeatingthem:theperceptualsimilarityof
nasalconsonantsmayhaveinterferedwiththeensuingproduction.
117
4.3. FutureDirections
Themostobviousfollow-uptobothStudies1and2arereplications.InStudy1,
replicationsshouldincludeagreaternumberoftalkers,aslistenersappearedsensitiveto
fine-grainedphoneticdifferencesbetweentalkers(seedifferencesbetweenExperiments
1Aand1B).InStudy2,replicationsshouldincludeagreaternumberofparticipants,witha
designmodifiedinanumberofways(lessphoneticallysimilarvowelsinVowelcondition;
lesssimilarexperimentallyconstrainedconsonants;etc.).AnalternatedesignforStudy2
wouldexposelistenerstoa1st-orderconstraintinaninitialblock,followedbyareversalin
thefollowingblock,withmodeltalkeridentitychangingineachblock(Anderson,etal.,
2019).Whilepreviousstudieshaveshownreducedlearningforthereversedconstraint,if
speakersaresensitivetolanguagebackgroundinadaptationinproductiontheymaydetect
acontextchangefromoneblocktotheother,andadapttoanequaldegreeafterthe
reversal.
WehypothesizedthatlearnersarerecruitingtheirL2acquisitionfacultiesin
phonotacticadaptation,andtreatingthelaboratoryexposureasanovellanguage.
Accordingtoouraccount,thismotivatesadaptation,aslearnersareabletoseparatetheir
experiencewiththeirnativelanguagefromthenovelexperimentalinput,andtherefore
rapidlyadapttonovelconstraints.Iflearnersbelievetheyarebeingexposedtospeech
fromtheirnativelanguage,however,theyshouldbelesslikelytoadapt,astheyhave
extremelystrongpriorsabouttheirnativelanguagephonotacticconstraintsfromalifetime
ofexperience.Wepredictthatminutesoflaboratoryexposureinthelearner’sL1willnot
118
beenoughtoovercomethesepriors.Onewaytotestthispredictionwouldbetoexpose
learnerstorealwordsintheirnativelanguage,presentedinsemanticallymeaningful
contexts.Forexample,atonguetwistermadeupofrealwords(e.g.,pothamkingfin)could
beinitiallypresentedtoparticipantsaspictures,beforeorthographicsupportappears(just
astwisterswereinitiallypresentedauditorilyinStudy2).Ifrapidphonotacticadaptationis
partoftheprocessofL2acquisitionwewouldnotexpectittobeactiveinnativespeech
contexts;subsequently,wewouldpredictlowratesofadaptationinsuchcases.
IfphonotacticadaptationisindeedtiedtoL2acquisition,wemightalsoexpectthat
participants’performanceinthesetasksmaypredictoutcomesforlongertermL2learning.
Forexample,ifrapidphonotacticadaptationisoneofthekeystounlockingword
segmentationearlyinL2acquisition,participantswhoadapttoagreaterdegree,andare
thereforecapableofquicklylearningcuestonovelwordboundaries,mayshowgreater
abilitytosegmentwordsintheirL2afterexposure.Astudyalongtheselinescouldgive
participantsapre-testusingthemethodsinStudy1beforeanL2immersionprogram,
followedbyapost-testwordsegmentationtask(e.g.,McQueen,1998).
Anotherquestionthatarisesfromthisresearchishowpresentingexplicit
informationabouttalkerswouldaffectparticipantbehavior.Inspeechperceptionresearch,
modifyinglistenerexpectationsabouttalkercharacteristicssuchasdialect,eveninsubtle
ways,canhaveimportantconsequencesforspeechperception(e.g.,Hay&Drager,2010).
Evenmanipulatingthenumberofspeakersthatlistenersexpecttohearcanaffect
processingofthesamelinguisticinput(MagnusonandNusbaum,2007).Inthecurrent
studies,wehypothesizedthatadaptationisinducedbythelearner’sbeliefthattheyare
119
beingexposedtodifferentlanguages.Thisbeliefabouttalkers’languagebackgrounds
comesfromphoneticsalone,asparticipantsaregivennoexplicitinformationaboutthe
experimentaltalkerswhatsoever.Ifwepresentedexplicitinformationthatspeakers
differedintheirlanguagebackgrounds(e.g.,“BarbaragrewupspeakingEnglishinOhio,
whileBélagrewupspeakingHungarianinBudapest”)wemightexpectittostrengthen
participants’confidencethatthetalkersdonotsharealanguagebackground,andthus
boostadaptation.Alternatively,presentingexplicitinformationthattalkerssharea
languagebackgroundmaydampenadaptation.
Insomeconditions,itmaybethecasethatlistenersarealreadyfullyconfidentin
theirbeliefsabouttalkers’languagebackgrounds,suggestingthattheirconfidencecould
notbeincreasedfurtherbytop-downinformation.IntheDifferentconditioninStudy1,for
example,listenersmayhavealreadybeenfullyconfidentthattalkersdifferedinthose
cases,asmodulatingthe“non-nativeness”ofthephoneticvowelcuesintheStrongvs.
WeakDifferentconditionsdidnotchangethedegreeofadaptation.Inthiscase,explicit
informationthattalkersdifferedintheirlanguagebackgroundsmaynotaffectadaptation.
Informationthattalkerssharealanguagebackground,however,maydecreaseadaptation.
IntheSharedconditions,listenersappearedtoshowvaryingdegreesofconfidenceabout
thelanguagebackgroundsoftalkers.WemightexpectintheNon-NativeSharedconditions,
forexample,thatiflistenersareexplicitlytoldthetwospeakerssharealanguage
background,thedegreeofadaptationmightdecrease.Inthissamecontext,pushing
participantsinthereversedirection,bygivingtheminformationthatthetwotalkersdiffer
120
intheirlanguagebackgrounds,mightoverridethephoneticsimilaritiesofthetwotalkers
andincreaseadaptation.
Anothernaturalextensionofthisdissertation,givenourfindingsthatlearnersadapt
toagreaterdegreewhentalkersdifferintheirlanguagebackgrounds,iswhetherlearners
wouldalsoshowincreasedadaptationtotalkersofdifferentdialectsofthesamelanguage
(e.g.,tworegionaldialectsoftheUnitedStates).Theexactnatureoflistener’sprior
experiencewithdialectsandphonotacticvariationisnotentirelyclear.Ifdialectsvaryin
theirphonotactics,buttoalesserextentthanlanguagesvary,wewouldpredictlisteners
wouldadapttoamoderatedegreetotalkersofdifferentdialects.Whiledialectsalmost
certainlyhavelesspervasivevariationincategoricalconstraints(e.g.,there’snodialectof
Englishthatallows[ŋ]inonsetposition),theymayshowmorevariationforgradient
phonotacticconstraints(e.g.,[s]appearsmoreofteninonsetthan[z]).
Whilequantifyingphonotacticvariationisakeychallengeforfuturework,itis
unclearhowtooperationalizeandcomparedistancebetweenphonotacticgrammars.
Whileweareconfidentthatlanguagesdiffertoagreaterdegreethanindividualtalkers,
thereisnostraightforwardmethodtodrawingeitherquantitativeorqualitative
distinctionsbetweenpairsoflanguagesforanumberofreasons.First,itmaybethecase
thatnotallphonotacticdifferencesarecreatedequal.Differencesthathaveawiderimpact
onthelexiconofalanguage,suchasconstraintsontheshapeofsyllables(e.g.,languages
limitedtoonlyCVsyllablesvs.languagesthatallowmorecomplexsyllables)shouldbe
weightedmoreheavilythandifferenceswithregardstosinglesounds(e.g.,languagesthat
allow[ŋ]inonsetvs.thosethatdonot).Second,itisunclearhowtocomparelanguagewith
121
widelydifferentsoundinventories.Unlike,forexample,Cantonese,Englishdoesnothave
phonotacticconstraintsonlexicaltone,butofcourselexicaltoneisnotafeatureofEnglish
atall.Finally,itisunclearifphonotacticdistanceshouldbequantifiedinan“objective”
way,orifdistanceshouldbemeasuredbyhowdetectablephonotacticdifferencesare
betweenlanguagesfromthespeaker’spointofview.Thisperceiveddistancemaybethe
mostimportantaspectforthepurposesofadaptation.
Finally,wecouldaskwhatotherdomainsthislinkbetweentalkerlanguage
backgroundandunderlyinggrammarextendto.Speakersexperiencevariationatevery
leveloflinguisticrepresentation.Manyofthesedomainsmayholdasimilarstructurein
variationtophonotactics—ahighdegreeofvariationbetweentalkersofdifferentlanguage
varieties,andalowdegreeofvariationbetweentalkersofthesamelanguagevariety.As
such,wewouldexpectthesameprinciplesofcausalinferencetoapply.Forexample,this
inferencemayextendtolearningofnovelorunlikelysyntacticormorphologicalstructures.
Inthecaseofartificiallanguageparadigms(e.g.,Schumacher,Pierrehumbert,&Lashell,
2014),ifstimuliarepresentedbynon-nativetalkers,itmayboostadaptation.For
adaptationinnativelanguagecontexts(e.g.,Jaeger&Snider,2013),ifstimuliarepresented
bytalkersofdifferentdialectsitmayalsoincreaseadaptation,assyntaxmayvarytoa
greaterdegreebetweenspeakersofdifferentdialects(e.g.,Labov,1969)thanitdoes
betweenindividualswithinaspeechcommunity.
122
4.4. Conclusions
Intwostudies,wehaveinvestigatedthewaysinwhichspeakers’priorexperience
guidesphonotacticadaptation.Wehypothesizedthatspeakersadapttorelevantor
systematicvariation,whileignoringirrelevantorincidentalvariationforthetaskathand,
basedontheirpreviousexposuretophonotacticvariation.Wefoundstrongevidencefor
thiseffectinperception,butonlyweak/inconclusiveevidenceinproduction.Thisevidence
extendstheoriesofadaptationinwhichspeakersmakeinferencesaboutthecausesof
variationtoanoveldomain.Inaddition,wereframethephenomenonofphonotactic
learningasapartoftheL2acquisitionfaculty.Asawhole,thisdissertationexploreshow
speakerscontendwithandadapttoendemicvariation,sheddinglightonthemechanisms
andmotivationsforadaptation.
123
5. Appendix
5.1. AppendixA–Study1vowelacousticsanalysis
ResultsfromExperiments1Aand1BfoundnodifferencebetweenWeakandStrong
Differentconditions,whichweredifferentiatedbyeitherarelativelyweakphoneticcueto
theFrenchtalkers’non-nativelanguagebackground(useoftheFrenchvowel[u])ora
strongercue(useoftheFrenchvowel[y]).Oneexplanationforthislackofeffectisthatthe
Weakconditionincludedasufficientnumberofacousticcuessuchthatlistenerscould
confidentlyinferthattheFrenchtalkerswerenon-native.Analternative,however,isthat
theFrenchtalkers’productionsof[u]and[y]werenotsufficientlyacousticallydistinct,and
thereforethevowelmanipulationmadelittledifferencetolisteners.Toinvestigatethis
possibility,apost-hocacousticanalysiswascompletedoftheFrenchspeakers’vowel
productionsinExperiments1Aand1B.
Vowelsintervalsforeachitemwerehand-markedinPraat.F1andF2valueswere
thenautomaticallymeasuredatthemid-pointofeachvowelusingascript.Atotalof324
vowelsweremeasured(108foreachoftheFrenchtalkers).NotethatdifferentLPC(linear
predictivecoding)settingswereusedtoaccuratecapturemaleandfemaleformantvalues.
124
Figure5.1.ScatterplotofFrenchtalkers’firstandsecondvowelformants.Eachpointisavowel,withvowelidentityindicatedbytheappropriateIPAsymbol.Colorindicatestalkerdifferences.
AsshowninFigure5.1,[y]and[u]areacousticallydistinctbasedonthedifferencesin
F1andF2forthefemaleFrenchspeakers.Thissuggeststhelackofaneffectbasedoncue
strengthdoesnotstemfromalackofacousticdifferencesacrossconditions.Further
evidencethatEnglishlistenerscandistinguishthese[y]and[u]stimulicomesfromSteele,
etal.(2015).UsingthesamestimuliasExperiment1A,theyfoundthatnativeEnglish
listenerscanacquirephonotacticconstraintsconditionedon[y]vs.[u]–suggestingthe
stimuliareperceptuallydistinct.
125
5.2. AppendixB–Study1pilotstudy
PriortoExperiment1,apilotstudywasrunthatincluded3ofthe4conditions
includedinExperiment1A:NativeShared(2Englishtalkers),WeakDifferent(1English
talkerand1Frenchtalkerwiththe[u]vowel),andStrongDifferent(1Englishtalkerand1
Frenchtalkerwitheh[y]vowel).OtherthantheexclusionoftheNon-nativeShared
condition,thestimuli,design,andprocedureofthepilotwereidenticaltothoseof
Experiment1A.ThepoweranalysisforExperiment1(seeAppendixC)wasbasedonthe
resultsfromthispilot.
Participants
Toreachthetargetof48participants(16percondition)whopassedthe
experimentalcriteria,atotalof85nativespeakersofEnglishwererecruitedonAMT
(passingrateof56.5%).
DataAnalysis
DataanalysiswasidenticaltothatinExperiment1Awithoneexception:theaccent
term,comparingthetwoSharedconditions,wasnotincluded,astherewasonlyoneshared
condition.Thepilotdatawasanalyzedusingalogisticmixed-effectsregression,with
participantresponsesasthedependentmeasure.Fixedeffectsincludedlegality,andtwo
contrast-codedterms:languagedifference(i.e.,Sharedvs.Differentconditions),and
strength(i.e.,Weakvs.StrongDifferentconditions).Interactiontermswereincluded
126
betweenlegalityandbothcontrast-codedterms.Randomeffectsincludedrandom
interceptsandrandomslopesbylegalityforbothparticipantsanditems.Inaddition,
follow-upanalyseswererunonindividualconditions,whichincludedafixedeffectof
legality,andrandomintercepts,aswellasslopesbylegalityforitems(themodelsdidnot
convergewithrandomslopesbyparticipant).
PilotResults
Theanalysisrevealedamaineffectoflegality(β=0.49,s.e.β=0.12,χ2(1)=15.23,p
<0.001),suggestingthatparticipantswere,overall,abletolearntheconstraint.Inaddition,
therewasamarginalinteractionbetweenthesharedtermandlegality(β=-0.63,s.e.β=
0.34,χ2(1)=3.31,p=0.07),providingweakevidencethatparticipantsadaptedtoagreater
degreeintheDifferentconditions.Thiswasconsistentwithafollow-upanalysisshowing
thattherewasasignificantdifferencebetweenparticipantresponsesonlegalandillegal
syllablesinbothoftheDifferentconditions(Strong:β=0.73,s.e.β=0.19,χ2(1)=13.1,p<
0.001;Weak:β=0.46,s.e.β=0.19,χ2(1)=6.14,p<0.05),butnosuchdifferenceinthe
SharedNativecondition(β=0.19,s.e.β=0.20,χ2(1)=0.92,p=0.34).Thissuggestedthat
listenersonlyadaptedtotalker-specficphonotacticconstraintsifspeakersdifferedintheir
languagebackground,aswasfoundinExperiments1and2.Therewasnosignificant
interactionbetweenstrengthandlegality(β=0.32,s.e.β=0.30,χ2(1)=1.16,p=0.28),
suggestingbothconditionsprovidedsufficientcuesforlistenerstoidentifyadifferencein
languagebackground,similartoExperiment1.
127
5.3. AppendixC–Study1poweranalysis
Thenumberofparticipantswassettoyieldsufficientstatisticalpower(β>.8).
PowerwasestimatedbyMonteCarlosimulationsbasedonresultsfromapilotstudy(see
AppendixBfordetails).Usingtheestimatesforeachfixedandrandomeffectinthelogistic
mixedeffectsmodelfittothesepilotdata,wegenerated1000simulateddatasets.Foreach
simulateddataset,werandomlyandindependentlysampledeachfixedeffectvaluefroma
normaldistribution(withthemeansettotherespectivecoefficientestimateandstandard
deviationsettothecorrespondingstandarderrorestimate)andindependentlysampled
eachrandomeffectbasedontheestimatedrandomeffectdistributions(correlations
betweencoefficientswerenotincorporatedintooursamplingprocedure).Thesewerethen
usedtogenerateasetofrecognitionmemorytestresponse.Wethenfitthesame
regressionmodeltothesesimulatedresponses.(Ifthemodelfailedtoconverge,we
generatedanewsimulateddataset.)Statisticalpowerβwasestimatedbytheproportion
ofthe1000modelsinwhichthecrucialinteractionterm—betweenthefixedeffectsfor
legalityandshared/differentlanguagebackground—wasfoundtobesignificant.We
increasedthenumberofparticipantsiteratively,generatingnovelsimulateddatasetsand
runningnewmodelswitheachiteration,untilwereachedthethresholdofβ>.8.This
thresholdwasreachedwith64participantspercondition(estimatedβ=.804).
128
5.4. AppendixD–Study1passingrates
RecallthatexperimentalcriteriainExperiments1Aand1Bwereasfollows:inthe
generalizationphase,participantshadtocorrectlyacceptatleast90%ofpreviouslyheard
items(i.e.,ahitrateofover90%)andcorrectlyrejectatleast10%ofnovelitems.These
criteriaensuredthatparticipantswereabletorecallitemstheyhadpreviouslyheard
multipletimes,andthattheywereabletodifferentiatebetweenpreviouslyhearditems
andnovelitems.InExperiment2,thesecriteriawereloosenedduetothephonological
confusabilityofthestimulusset:thehitratecriterionwasloweredfrom90%to85%.A
thirdcriterionwasalsoadded:listeners’hitratecouldnotexceedtheirfalsealarmrate,to
ensurethatlistenerscoulddifferentiatenovelandfamiliaritems.Forexample,a
participantwithahitrateof85%mustcorrectlyrejectatleast15%ofnovelitems(i.e.,a
falsealarmratenohigherthan85%).
129
Table5.1.PassingratesforeachconditionandexperimentinStudy1.
Experiment Condition Totalparticipants
PassingParticipants
PassingRate
1A NativeShared 124 64 51.6%1A Non-NativeShared 101 66 65.3%1A WeakDifferent 118 66 55.9%1A StrongDifferent 112 64 57.1%1B Non-NativeShared 158 66 41.8%1B WeakDifferent 119 64 53.8%1B StrongDifferent 141 64 45.4%2 Non-NativeShared 160 66 41.3%2 Non-NativeDifferent 157 67 42.7%2 MixedDifferent 124 69 55.6%n.b.Thesamenumberofparticipants(64)wereanalyzedineachcondition.Thenumberofpassingparticipantssometimesexceededthisduetotechnicallimitationsinourexperimentalpipeline. AsshowninTable5.1,passingratesrangedfrom41.3%to65.3%between
conditions.Overall,655participantspassedthecriteriaoutof1314participants(49.8%).
Thiswasinlinewithpreviousresultsusingthisparadigm(Denby,etal.,2018).Inapost-
hocanalysis,weinvestigatetherelationshipbetweenthehitrateinthegeneralization
phaseandthelegalityeffect.AsFigure5.2shows,participantswithahitratelowerthan
roughly75%showlittletonolegalityeffect.Thisisunsurprising:if,forexample,a
participantcorrectlyacceptsonlyhalfoffamiliaritems,theyaresimplyatchance,and
thereforewillnotshowanydifferencesbetweenlegalandillegalgeneralizationitems,as
theyarelikelyguessing.Participantswhosehitrateismuchlowerthan50%mayhave
misinterpretedtheexperimentalinstructions,andsimplyanswered“no”toanyitemsthey
hadnotencounteredpriortotheexperiment(ratherthanwithintheexperiment).
130
Figure5.2.Scatterplotofhitrate(%yesonfamiliaritems)ingeneralizationphasebylegalityadvantage(falsealarmrateforlegalitemsminusfalsealarmrateonillegalitems)forallexperimentsinStudy1.Eachdotrepresentsasingleparticipant;colorsrepresentwhetherparticipantspassedorfailedcriteria.LinesrepresentLoessregression;shadingrepresents95%confidenceinterval
Notethattherelationshipbetweenhitrateandlegalityadvantageisnon-linear—for
participantswhofailthecriteria,thelegalityadvantagepeaksaroundahitrateof85%;as
thehitrateincreasesto100%,thelegalityadvantagefallsbackdowntoalmost0%.Thisis
alsoanexpectedresult,assuchparticipantsarefailingthecriteriabasedonahighfalse
alarmrate:theyareresponding“yes”toalmosteveryitem,regardlessofwhetheritis
familiarornovel.Detectingdifferencesinresponsepatternsbetweenlegalandillegalnovel
itemsisessentiallyimpossiblewithsuchahighoverallfalsealarmrate.
131
Theseresultssuggestthatthecriteriawerenecessarytofilteroutparticipantswho
werebiasedtowardsalwaysresponding“yes”oralwaysresponding“no”,aswellasthose
whoansweredrandomly(i.e.,ahitrateof~50%).Italsoappearsasthoughthecriteria
mayhavebeenslightlytoorestrictive,asparticipantswhosehitratewasaboveroughly
75%appearedtobetrackingtheconstraint,asshownbytheirincreasedlegality
advantage.Inasecondpost-hocanalysis,were-plottedthedatawhilelooseningthe
criteriatoincludeparticipantswithahitrateaslowas75%.AsinExperiment2,we
includedacriterionthatparticipants’falsealarmratemustbelowerthantheirhitrate,to
ensuretheyareabletodifferentiatefamiliarandnovelitems.
Looseningthecriteriaresultedinanadditional152participantspassing,increasing
theoverallpassingratefrom49.8%to61.4%.AscanbeseeninFigures5.3–5.5,the
resultsoftheexperimentdonotqualitativelychangewiththeadditionofthese
participants.Thissuggeststhecriteriaasoriginallysetweresomewhatoverlyrestrictive,
withanadditional~10%ofparticipantsunnecessarilyexcluded.Basedontheseresults,we
recommendthatfutureexperimentswithsimilardesignsshouldloosenthecriteriato75%.
132
Figure5.3.LegalityadvantageforExperiment1A,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval.
Figure5.4.LegalityadvantageforExperiment1B,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval.
133
Figure5.5.LegalityadvantageforExperiment2,withhitratecriterionloweredto75%.Errorbarsreflectbootstrapped95%confidenceinterval.
134
5.5. AppendixE–Study1modelresults
Table5.2.Experiment1AFixedEffects.
FixedEffect Coefficient StandardError χ2(1) P
(Intercept) 0.30** 0.09 Legality 0.64*** 0.06 73.63 <0.0001LanguageDifference 0.00 0.26 0.00 0.99Strength -0.28 0.18 2.43 0.12Accent -0.34 0.18 3.41 0.06Legality:LanguageDifference 0.72*** 0.20 12.59 <0.001Legality:Strength -0.05 0.14 0.15 0.70Legality:Accent 0.45** 0.14 9.78 <.01Table5.3.Experiment1Arandomeffects.
RandomEffects VarianceIntercept(item) 0.31Slope(itembylegality) 0.06Intercept(participant) 0.91Slope(participantbylegality) 0.06Table5.4.Experiment1Bfixedeffects.
FixedEffect Coefficient StandardError χ2(1) p
(Intercept) 0.52 0.09
Legality 1.02*** 0.07 127.64 <0.0001LanguageDifference -1.06*** 0.19 28.82 <0.0001Strength 0.16 0.16 0.99 0.32Legality:LanguageDifference 0.57*** 0.18 9.81 <0.01Legality:Strength -0.02 0.15 0.03 0.87
135
Table5.5.Experiment1Brandomeffects.
RandomEffects VarianceIntercept(Item) 0.36Slope(itembyLegality) 0.06Intercept(Participant) 0.67Slope(ParticipantbyLegality) 0.09Table5.6.Experiment2fixedeffects.
FixedEffect Coefficient StandardError χ2(1) P
(Intercept) 0.97*** 0.08 Legality 0.76*** 0.08 62.83 <0.0001LanguageDifference -0.20 0.19 1.13 0.29Native 0.14 0.19 0.52 0.47Legality:LanguageDifference 0.71*** 0.21 11.08 <0.001Legality:Native -0.04 0.21 0.03 0.87Table5.7.Experiment2randomeffects.
RandomEffects VarianceIntercept(item) 0.15Slope(itembyLegality) 0.08Intercept(participant) 0.45Slope(participantbyLegality) 0.09
136
5.6. AppendixF–Tonguetwistersamplesizeanalysis
Ananalysisof15previoustonguetwisterexperimentsforwhichdesigndetailsand
resultswereaccessiblefoundawiderangeoftotalnumberofsyllablesanalyzedanderror
rates(seeTable5.8).Theaveragenumberofitemsanalyzed(wordsorsyllables,depending
onthedesignoftheexperiment)was49,152(rangeof9,216-184,320).Thecurrent
experimentexposesparticipantstosecond-orderconstraints;theaveragenumberofitems
forpreviousexperimentswithsuchconstraintswas53,453(rangeof18,432-82,944).
137
Table5.8Numberofitemsanalyzed(wordsorsyllables),errors,errorrates,andconstraintorderforprevioustonguetwisterexperiments.
Paper Experiment TotalItems
TotalErrors
ErrorRate
Order
Delletal.(2000) 1 36,864 3065 8.3% 1stDelletal.(2000) 2 36,864 3584 9.7% 1stDelletal.(2000) 3 36,864 1769 4.8% 2ndGoldrick(2004) 1 184,320 6762 4.8% 1stKittredge&Dell(2016)
1 64,512 5010 5.7% 1st
Smalleetal.,(2017) 1-Adults 55,296 1240 2.2% 2ndTaylor&Houghton(2005)
1 36,864 1313 3.6% 1st
Taylor&Houghton(2005)
2 9,216 745 8.1% 1st
Taylor&Houghton(2005)
3 9,216 729 7.9% 1st
Warker(2013) 1 82,944 3926 4.7% 2ndWarker(2013) 2 55,296 3355 6.1% 2ndWarker&Dell(2006)
1a 18,432 500 2.7% 2nd
Warker&Dell(2006)
1b 18,432 1074 5.8% 2nd
Warker&Dell(2015)
1 55,296 4043 7.3% 1st
Warkeretal.(2008) 1 73,728 5460 7.4% 2ndWarkeretal.(2009) 1 18,432 391 2.1% 1stAninitialpilotexperimentwaspreviouslyconductedwith9,216totalitemspercondition
(36,864total).11ofthe32participantsessionswereanalyzedforerrors,foratotalof
12,672items.Atotalof433errorswerefound,foranerrorrateof3.4%.Thiserrorrate
waswithintherangeofpreviouslyfounderrorrates,althoughitwasonthelowendofthe
range.Thelowerrorrate,however,andtherelativelysmallnumberofitems,renderedthe
resultsuninterpretable,especiallywhensplitbycondition.Thecurrentexperiment
increasesthesamplesizeto110,592(27,648percondition;seeabovefordetails).
138
5.7. AppendixG–Study2modelresults
5.7.1. Within-conditionmodels;allsessions;[m]-[n]swapsincluded
Table5.9.Differentcondition;allsessions;[m]-[n]swapsincluded.
FixedEffect Coefficient StandardError χ2(1) P(Intercept) 0.32 0.14
Constraint -0.97*** 0.19 25.86 <0.001Session 0.28 0.18 2.34 0.13Constraint:Session 0.44 0.36 1.48 0.22
RandomEffects
VarianceTargetSyllable
0.46
Table5.10.NativeSharedcondition;allsessions;[m]-[n]swapsincluded.
FixedEffect Coefficient StandardError χ2(1) P(Intercept) 0.43 0.13
Constraint -0.92*** 0.17 30.74 <0.001Session 0.1 0.16 0.38 0.54Constraint:Session -0.61 0.31 3.91 0.05
RandomEffects
VarianceTargetSyllable
0.43
139
Table5.11.Non-NativeSharedcondition;allsessions;[m]-[n]swapsincluded.
FixedEffect Coefficient StandardError χ2(1) P(Intercept) 0.62 0.2
Constraint -1.64*** 0.21 64.2 <0.001Session -0.27 0.18 2.25 0.13Constraint:Session 0.42 0.36 1.36 0.24
RandomEffects
VarianceTargetSyllable
1.2
Table5.12.Vowelcondition;allsessions;[m]-[n]swapsincluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 0.94 0.18
Constraint -1.47*** 0.28 26.9 <0.001Session 0.11 0.27 0.16 0.69Constraint:Session 0.66 0.54 1.51 0.22
RandomEffects
VarianceTargetSyllable
0.86
5.7.2. Within-conditionmodels;[m]-[n]swapsexcluded
Table5.13.Differentcondition;allsessions;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.12 0.16
Constraint 0.49 0.26 3.8 0.05Session 0.27 0.25 1.21 0.27Constraint:Session 0.39 0.49 0.63 0.43
RandomEffects
VarianceTargetSyllable
0.34
140
Table5.14.NativeSharedcondition;allsessions;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.17 0.14
Constraint 0.38 0.21 3.33 0.07Session -0.06 0.21 0.09 0.76Constraint:Session -0.78 0.41 3.58 0.06
RandomEffects
VarianceTargetSyllable
0.26
Table5.15.Non-NativeSharedcondition;allsessions;[m]-[n]swapsexcluded
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.52 0.17
Constraint 0.01 0.25 0 0.97Session -0.3 0.23 1.68 0.2Constraint:Session 0.38 0.45 0.69 0.41
RandomEffects
VarianceTargetSyllable
0.42
Table5.16.Vowelcondition;allsessions;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.81 0.25
Constraint 0.13 0.38 0.11 0.74Session 0.19 0.37 0.28 0.59Constraint:Session 0.8 0.73 1.22 0.27
RandomEffects
VarianceTargetSyllable
0.81
5.7.3. Within-conditionmodels;session1excluded;[m]-[n]swapsexcluded
141
Table5.17.Differentcondition;session1excluded;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.23 0.2
Constraint 0.73* 0.35 4.76 0.03
RandomEffects
VarianceTargetSyllable
0.81
Table5.18.NativeShared;session1excluded;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError χ2(1) p(Intercept) 1.23 0.2
Constraint -0.01 0.31 0 0.98
RandomEffects
VarianceTargetSyllable
0.54
Table5.19.Non-NativeSharedcondition;session1excluded;[m]-[n]swapsexcluded.
FixedEffect Coefficient StandardError X2(1) p(Intercept) 1.39 0.2
Constraint 0.14 0.33 0.18 0.67
RandomEffects
VarianceTargetSyllable
0.38
Table5.20.Vowelcondition;session1excluded;[m]-[n]swapsexcluded
FixedEffect Coefficient StandardError X2(1) p(Intercept) 2.13 0.42
Constraint 0.62 0.61 1.13 0.29
RandomEffects
VarianceTargetSyllable
1.57
142
6. References
Anderson,N.D.,Holmes,E.W.,Dell,G.S.,&Middleton,E.L.(2019).Reversalshiftin
phonotacticlearningduringlanguageproduction:Evidenceforincrementallearning.
JournalofMemoryandLanguage,106(March2018),135–149.
https://doi.org/10.1016/j.jml.2019.03.002
Atagi,E.,&Bent,T.(2013).Auditoryfreeclassificationofnonnativespeech.Journalof
Phonetics,41,509–519.https://doi.org/10.1016/j.wocn.2013.09.003
Baese-Berk,M.M.,Bradlow,A.R.,&Wright,B.a.(2013).Accent-independentadaptationto
foreignaccentedspeech.TheJournaloftheAcousticalSocietyofAmerica,133(3),
EL174-80.https://doi.org/10.1121/1.4789864
Barr,D.J.,Levy,R.,Scheepers,C.,&Tily,H.J.(2013).Randomeffectsstructurefor
confirmatoryhypothesistesting:Keepitmaximal.JournalofMemoryandLanguage,
44(0),255–278.https://doi.org/http://doi.org/10.1016/j.jml.2012.11.001
Bernard,A.(2015).Anonsetisanonset:Evidencefromabstractionofnewly-learned
phonotacticconstraints.JournalofMemoryandLanguage,78(2015),18–32.
https://doi.org/10.1016/j.jml.2014.09.001
Bernard,A.(2017).Novelphonotacticlearning:Trackingsyllable-positionandco-
occurrenceconstraints.JournalofMemoryandLanguage,96(2017),138–154.
https://doi.org/10.1016/j.jml.2017.05.006
Best,C.T.,McRoberts,G.W.,&Goodell,E.(2001).Discriminationofnon-nativeconsonant
contrastsvaryinginperceptualassimilationtothelistener’snativephonological
143
system.TheJournaloftheAcousticalSocietyofAmerica,109(2),775–794.
https://doi.org/10.1121/1.1332378
Blom,E.,Küntay,A.C.,Messer,M.,Verhagen,J.,&Leseman,P.(2014).Journalof
ExperimentalChildThebenefitsofbeingbilingual :Workingmemoryinbilingual
Turkish–Dutchchildren.JournalofExperimentalChildPsychology,128,105–119.
https://doi.org/10.1016/j.jecp.2014.06.007
Bosch,L.,&Sebastián-Gallés,N.(2001).EvidenceofEarlyLanguageDiscrimination
AbilitiesinInfantsfromBilingualEnvironments.Infancy,2(1),29–49.
https://doi.org/10.1207/S15327078IN0201_3
Bradlow,A.R.,&Bent,T.(2008).Perceptualadaptationtonon-nativespeech.Cognition,
106(2),707–729.https://doi.org/10.1016/j.cognition.2007.04.005
Brent,M.R.,&Cartwright,T.a.(1996).Distributionalregularityandphonotactic
constraintsareusefulforsegmentation.Cognition,61(1–2),93–125.
https://doi.org/10.1016/S0010-0277(96)00719-6
Buhrmester,M.,Kwang,T.,&Gosling,S.D.(2011).Amazon’smechanicalTurk:Anew
sourceofinexpensive,yethigh-quality,data?PerspectivesonPsychologicalScience,
6(1),3–5.https://doi.org/10.1177/1745691610393980
Caramazza,A.,&Yeni-Komshian,G.H.(1974).VoiceonsettimeintwoFrench
dialects.Journalofphonetics,2(3),239-245.
Carlson,M.T.,Goldrick,M.,Blasingame,M.,&Fink,A.(2016).Navigatingconflicting
phonotacticconstraintsinbilingualspeechperception.Bilingualism,19(5),939–954.
https://doi.org/10.1017/S1366728915000334
144
Chomsky,N.,&Halle,M.(1965).Somecontroversialquestionsinphonologicaltheory.
JournalofLinguistics,1(2),97–138.https://doi.org/10.1017/S0022226700001134
Creel,S.C.,Aslin,R.N.,&Tanenhaus,M.K.(2008).Heedingthevoiceofexperience:therole
oftalkervariationinlexicalaccess.Cognition,106(2),633–664.
https://doi.org/10.1016/j.cognition.2007.03.013
Dart,S.N.(1998).ComparingFrenchandEnglishcoronalconsonantarticulation.Journalof
Phonetics,26(1),71–94.https://doi.org/10.1006/jpho.1997.0060
Dell,GS,Reed,K.D.,Adams,D.R.,&Meyer,A.S.(2000).Speecherrors,phonotactic
constraints,andimplicitlearning:astudyoftheroleofexperienceinlanguage
production.JExpPsycholLearnMemCogn,26(6),1355–1367.
https://doi.org/10.1037/0278-7393.26.6.1355
Dell,GaryS,&Chang,F.(2014).TheP-chain:relatingsentenceproductionanditsdisorders
tocomprehensionandacquisition.PhilosophicalTransactionsoftheRoyalSocietyof
London.SeriesB,BiologicalSciences,369(1634),20120394.
https://doi.org/10.1098/rstb.2012.0394
Denby,T.,Schecter,J.,Arn,S.,Dimov,S.,&Goldrick,M.(2018).Contextualvariabilityand
exemplarstrengthinphonotacticlearning.JournalofExperimentalPsychology:
Learning,Memory,andCognition,44(2),280–294.
https://doi.org/10.1037/xlm0000465
Drager,K.(2010).Sociophoneticvariationinspeechperception.LinguisticsandLanguage
Compass,4(7),473–480.https://doi.org/10.1111/j.1749-818X.2010.00210.x
Dupoux,E.,Kakehi,K.,Hirose,Y.,Pallier,C.,&Mehler,J.(1999).Epentheticvowelsin
145
Japanese:Aperceptualillusion?JournalofExperimentalPsychology:HumanPerception
andPerformance,25(6),1568–1578.https://doi.org/10.1037/0096-1523.25.6.1568
Eisner,F.,&McQueen,J.M.(2005).Thespecificityofperceptuallearninginspeech
processing.Perception&Psychophysics,67(2),224–238.
https://doi.org/10.3758/BF03206487
Flege,J.E.(1987).Theproductionof“new”and“similar”phonesinaforeignlanguage:
evidencefortheeffectofequivalenceclassification.JournalofPhonetics,(15),47–65.
Retrievedfromhttp://jimflege.com/files/Flege_new_similar_JP_1987.pdf
Fromkin,V.A.(1971).TheNon-AnomalousNatureofAnomalousUtterances.Language,
47(1),27–52.https://doi.org/10.2307/412187
Giles,H.(1973).AccentMobility:AModelandSomeData.AnthropologicalLinguistics,
15(2),87–105.https://doi.org/10.1111/j.1460-2466.2008.00398.x
Goldinger,S.D.(1996).Wordsandvoices:episodictracesinspokenwordidentification
andrecognitionmemory.JournalofExperimentalPsychology.Learning,Memory,and
Cognition,22(5),1166–1183.Retrievedfrom
http://www.ncbi.nlm.nih.gov/pubmed/8926483
Goldinger,S.D.(1998).Echoesofechoes?Anepisodictheoryoflexicalaccess.Psychological
Review,105(2),251–279.Retrievedfrom
http://www.ncbi.nlm.nih.gov/pubmed/9577239
Goldrick,M.(2004).Phonologicalfeaturesandphonotacticconstraintsinspeech
production.JournalofMemoryandLanguage,51,586–603.
https://doi.org/10.1016/j.jml.2004.07.004
146
Goldrick,M.(2011).LinkingSpeechErrorsandGenerativePhonologicalTheory.Language
andLinguisticsCompass,5(6),397–412.https://doi.org/10.1111/j.1749-
818X.2011.00282.x
Hay,J.,&Drager,K.(2010).Stuffedtoysandspeechperception.Linguistics,48(4),1689–
1699.https://doi.org/10.1515/ling.2010.027
Hillenbrand,J.,Getty,L.a,Clark,M.J.,&Wheeler,K.(1995).Acousticcharacteristicsof
AmericanEnglishvowels.TheJournaloftheAcousticalSocietyofAmerica,97(5Pt1),
3099–3111.https://doi.org/10.1121/1.411872
Jaeger,T.F.,&Snider,N.E.(2013).Alignmentasaconsequenceofexpectationadaptation:
Syntacticprimingisaffectedbytheprime’spredictionerrorgivenbothpriorand
recentexperience.Cognition,127(1),57–83.
https://doi.org/10.1016/j.cognition.2012.10.013
Kessler,B.,&Treiman,R.(1997).SyllableStructureandtheDistributionofPhonemesin
EnglishSyllables.JournalofMemoryandLanguage,37,295–311.
Kittredge,A.K.,&Dell,G.S.(2016).Learningtospeakbylistening:Transferofphonotactics
fromperceptiontoproduction.JournalofMemoryandLanguage,89(2016),8–22.
https://doi.org/10.1016/j.jml.2015.08.001
Kleinschmidt,D.F.(2018).Structureintalkervariability:Howmuchisthereandhowmuch
canithelp?Language,CognitionandNeuroscience,3798.
https://doi.org/10.1080/23273798.2018.1500698
Kleinschmidt,D.F.,&Jaeger,T.F.(2015).Robustspeechperception :Recognizethefamiliar
,generalizetothesimilar,andadapttothenovel.PsychologicalReview,148–203.
147
https://doi.org/10.1037/a0038695
Kraljic,T.,&Samuel,A.G.(2007).Perceptualadjustmentstomultiplespeakers.Journalof
MemoryandLanguage,56(1),1–15.https://doi.org/10.1016/j.jml.2006.07.010
Kraljic,T.,&Samuel,A.G.(2011).Perceptuallearningevidenceforcontextually-specific
representations.Cognition,121(3),459–465.
https://doi.org/10.1016/j.cognition.2011.08.015.Perceptual
Kraljic,T.,Samuel,A.G.,&Brennan,S.E.(2008).FirstImpressionsandLastResortsHow
ListenersAdjusttoSpeakerVariability,19(4),332–338.
Labov,W.(1969).Contraction,Deletion,andInherentVariabilityoftheEnglishCopula.
Language,45(4),715.https://doi.org/10.2307/412333
Levy,E.S.(2009).Ontheassimilation-discriminationrelationshipinAmericanEnglish
adults’Frenchvowellearning.TheJournaloftheAcousticalSocietyofAmerica,126(5),
2670–2682.https://doi.org/10.1121/1.3224715
Little,A.C.,&Apicella,C.L.(2016).FaceadaptationinanisolatedpopulationofAfrican
hunter-gatherers:Exposureinfluencesperceptionofother-ethnicityfacesmorethan
own-ethnicityfaces.PsychonomicBulletinandReview,23(2),439–444.
https://doi.org/10.3758/s13423-015-0919-z
Liu,L.,&Jaeger,T.F.(2018).Inferringcausesduringspeechperception.Cognition,
174(January),55–70.https://doi.org/10.1016/j.cognition.2018.01.003
Magnuson,J.S.,&Nusbaum,H.C.(2007).Acousticdifferences,listenerexpectations,and
theperceptualaccommodationoftalkervariability.JournalofExperimental
Psychology:HumanPerceptionandPerformance,33(2),391–409.
148
https://doi.org/10.1037/0096-1523.33.2.391
Marr,D.(1982).Vision.SanFrancisco:W.H.FreemanandCompany.Retrievedfrom
https://mechanism.ucsd.edu/teaching/f18/David_Marr_Vision_A_Computational_Inve
stigation_into_the_Human_Representation_and_Processing_of_Visual_Information.chap
ter1.pdf
Massaro,D.W.,&Cohen,M.M.(1983).Phonologicalcontextinspeechperception.
Perception&Psychophysics,34(4),338–348.https://doi.org/10.3758/BF03203046
Mattys,S.L.,&Jusczyk,P.W.(2001).Phonotacticcuesforsegmentationoffluentspeechby
infants.Cognition(Vol.78).https://doi.org/10.1016/S0010-0277(00)00109-8
McQueen,J.M.(1998).Segmentationofcontinuousspeechusingphonotactics.Journalof
MemoryandLanguage,39(1),21–46.https://doi.org/10.1006/jmla.1998.2568
Miller,G.A.,&Nicely,P.E.(1955).AnAnalysisofPerceptualConfusionsAmongSome
EnglishConsonants.TheJournaloftheAcousticalSocietyofAmerica,27(2),338–352.
https://doi.org/10.1121/1.1907526
Nooteboom,S.G.(1973).Thetongueslipsintopatterns.Speecherrorsaslinguisticevidence,
144-156.
Nygaard,L.C.,&Pisoni,D.B.(1998).Talker-specificlearninginspeechperception.
Perception&Psychophysics,60(3),355–376.https://doi.org/10.3758/BF03206860
Onishi,K.H.,Chambers,K.E.,&Fisher,C.(2002).Learningphonotacticconstraintsfrom
briefauditoryexperience.Cognition,83(1),13–23.https://doi.org/10.1016/S0010-
0277(01)00165-2
Otake,T.,Yoneyama,K.,Cutler,A.,&vanderLugt,A.(1996).Therepresentationof
149
Japanesemoraicnasals.TheJournaloftheAcousticalSocietyofAmerica,100(6),3831–
3842.https://doi.org/10.1121/1.417239
Pajak,B.,Fine,A.B.,Kleinschmidt,D.F.,&Jaeger,T.F.(2016).LearningAdditional
LanguagesasHierarchicalProbabilisticInference:InsightsFromFirstLanguage
Processing.LanguageLearning,1–45.https://doi.org/10.1111/lang.12168
Pardo,J.S.(2006).Onphoneticconvergenceduringconversationalinteraction.TheJournal
oftheAcousticalSocietyofAmerica,119(4),2382–2393.
https://doi.org/10.1121/1.2178720
Pardo,J.S.,Gibbons,R.,Suppes,A.,&Krauss,R.M.(2012).Phoneticconvergenceincollege
roommates.JournalofPhonetics,40(1),190–197.
https://doi.org/10.1016/j.wocn.2011.10.001
Pardo,J.S.,Urmanche,A.,Wilman,S.,&Wiener,J.(2017).Phoneticconvergenceacross
multiplemeasuresandmodeltalkers.Attention,Perception,andPsychophysics,79(2),
637–659.https://doi.org/10.3758/s13414-016-1226-0
Park,H.(2013).Detectingforeignaccentinmonosyllables:TheroleofL1phonotactics.
JournalofPhonetics,41(2),78–87.https://doi.org/10.1016/j.wocn.2012.11.001
Parlato-Oliveira,E.,Christophe,A.,Hirose,Y.,&Dupoux,E.(2010).Plasticityofillusory
vowelperceptioninBrazilian-Japanesebilinguals.TheJournaloftheAcousticalSociety
ofAmerica,127(6),3738–3748.https://doi.org/10.1121/1.3327792
Phatak,S.A.,&Allen,J.B.(2007).Consonantandvowelconfusionsinspeech-weighted
noise.TheJournaloftheAcousticalSocietyofAmerica,121(4),2312–2326.
https://doi.org/10.1121/1.2642397
150
Pickering,M.J.,&Garrod,S.(2004).Towardamechanisticpsychologyofdialogue.The
BehavioralandBrainSciences,27(2),169–190;discussion190-226.Retrievedfrom
http://www.ncbi.nlm.nih.gov/pubmed/15595235
Pickering,M.J.,&Garrod,S.(2007).Dopeopleuselanguageproductiontomake
predictionsduringcomprehension?TrendsinCognitiveSciences,11(3),105–110.
https://doi.org/10.1016/j.tics.2006.12.002
Pickering,M.J.,&Garrod,S.(2013).Anintegratedtheoryoflanguageproductionand
comprehension.BehavioralandBrainSciences,36(04),329–347.
https://doi.org/10.1017/s0140525x12001495
Pierrehumbert,J.(2001).Whyphonologicalconstraintsaresocoarse-grained.Language
andCognitiveProcesses,16(5–6),691–698.
https://doi.org/10.1080/01690960143000218
Pitt,M.a.,&McQueen,J.M.(1998).IsCompensationforCoarticulationMediatedbythe
Lexicon?JournalofMemoryandLanguage,39,347–370.
https://doi.org/10.1006/jmla.1998.2571
Qian,T.,Jaeger,T.F.,&Aslin,R.N.(2012).LearningtoRepresentaMulti-Context
Environment:MorethanDetectingChanges.FrontiersinPsychology,3(July),1–9.
https://doi.org/10.3389/fpsyg.2012.00228
Rebei,A.,Anderson,N.D.,&Dell,G.S.(2019).LearningthePhonotacticsofButtonPushing:
Consolidation,Retention,andSyllableStructure.JournalofExperimentalPsychology:
LearningMemoryandCognition.https://doi.org/10.1037/xlm0000686
Reinisch,E.,&Holt,L.L.(2014).Lexicallyguidedphoneticretuningofforeign-accented
151
speechanditsgeneralization.JournalofExperimentalPsychology:HumanPerception
andPerformance,40(2),539–555.https://doi.org/10.1037/a0034409
Richtsmeier,P.T.(2011).Word-types,notword-tokens,facilitateextractionofphonotactic
sequencesbyadults.LaboratoryPhonology,2(1).
https://doi.org/10.1515/labphon.2011.005
Schumacher,R.A.,Pierrehumbert,J.B.,&LaShell,P.(2014).ReconcilingInconsistencyin
EncodedMorphologicalDistinctionsinanArtificialLanguage.Proceedingsofthe36th
AnnualMeetingoftheCognitiveScienceSociety.Retrievedfrom
https://mindmodeling.org/cogsci2014/papers/500/paper500.pdf
Sidaras,S.K.,Alexander,J.E.D.,&Nygaard,L.C.(2009).Perceptuallearningofsystematic
variationinSpanish-accentedspeech.TheJournaloftheAcousticalSocietyofAmerica,
125(5),3306.https://doi.org/10.1121/1.3101452
Smalle,E.H.M.,Muylle,M.,Szmalec,A.,&Duyck,W.(2017).TheDifferentTimeCourseof
PhonotacticConstraintLearninginChildrenandAdults:EvidenceFromSpeechErrors.
JournalofExperimentalPsychology:Learning,Memory,andCognition,43(11),1821.
https://doi.org/10.1037/xlm0000405
StaumCasasanto,L.(2008).DoesSocialInformationInfluenceSentenceProcessing ?
Proceedingsofthe30thAnnualConferenceoftheCognitiveScienceSociety,(July2008),
799–804.Retrievedfrom
http://csjarchive.cogsci.rpi.edu/Proceedings/2008/pdfs/p799.pdf
Steele,A.,Denby,T.,Chan,C.,&Goldrick,M.(2015).LearningNon-NativePhonotactic
ConstraintsOvertheWeb.InTheScottishConsortiumforICPhS2015(Eds.)
152
ProceedingsoftheXVIIIthInternationalCongressofPhoneticSciences.
Storkel,H.L.,Armbrüster,J.,&Hogan,T.P.(2010).DifferentiatingPhonotacticProbability
andNeighborhoodDensityinAdultWordLearning,49(6),1175–1192.
https://doi.org/10.1044/1092-4388(2006/085).Differentiating
Strange,W.,Weber,A.,Levy,E.S.,Shafiro,V.,Hisagi,M.,&Nishi,K.(2007).Acoustic
variabilitywithinandacrossGerman,French,andAmericanEnglishvowels:Phonetic
contexteffects.TheJournaloftheAcousticalSocietyofAmerica,122(2),1111–1129.
https://doi.org/10.1121/1.2749716
Tomoschuk,B.,Ferreira,V.S.,&Gollan,T.H.(2018).Whenasevenisnotaseven:Self-
ratingsofbilinguallanguageproficiencydifferbetweenandwithinlanguage
populations.Bilingualism:LanguageandCognition,1–21.
https://doi.org/10.1017/S1366728918000421
Vitevitch,M.S.,&Luce,P.a.(1999).Probabilisticphonotacticsandneighborhoodactivation
inspokenwordrecognition.JournalofMemoryandLanguage,40(3),374–408.
https://doi.org/10.1006/jmla.1998.2618
Warker,J.a.(2013).InvestigatingtheRetentionandTimeCourseofPhonotacticConstraint
LearningFromProductionExperience.JournalofExperimentalPsychology:Learning,
Memory,andCognition,39(1),96–109.https://doi.org/10.1037/a0028648
Warker,J.A.,Xu,Y.,Dell,G.S.,&Fisher,C.(2009).Speecherrorsreflectthephonotactic
constraintsinrecentlyspokensyllables,butnotinrecentlyheardsyllables.Cognition,
112(1),81–96.https://doi.org/10.1016/j.cognition.2009.03.009
Warker,J.a,&Dell,G.S.(2006).Speecherrorsreflectnewlylearnedphonotactic
153
constraints.JournalofExperimentalPsychology.Learning,Memory,andCognition,
32(2),387–398.https://doi.org/10.1037/0278-7393.32.2.387
Warker,J.,Dell,G.S.,Whalen,C.,&Gereg,S.(2008).LimitsonLearningPhonotactic
ConstraintsfromRecentProductionExperience.JournalofExperimentalPsychology:
Learning,Memory,andCognition,34(5),1289–1295.https://doi.org/10.1007/s13398-
014-0173-7.2
Weber,A.,&Cutler,A.(2006).First-languagephonotacticsinsecond-languagelistening.
TheJournaloftheAcousticalSocietyofAmerica,119(1),597–607.
https://doi.org/10.1121/1.2141003
Webster,M.A.,Kaping,D.,Mizokami,Y.,&Duhamel,P.(2004).Adaptationtonaturalfacial
categories.Nature,428(6982),557–561.Retrievedfrom
http://dx.doi.org/10.1038/nature02420
Weiss,D.J.,Gerfen,C.,&Mitchel,A.D.(2009).SpeechSegmentationinaSimulatedBilingual
Environment:AChallengeforStatisticalLearning?LanguageLearningand
Development,5(1),30–49.https://doi.org/10.1080/15475440802340101
Wright,B.a,Baese-berk,M.,Marrone,N.,&Bradlow,A.R.(2016).Enhancingspeech
learningbycombiningtaskpracticewithperiodsofstimulusexposurewithout
practice,138(2),1–26.https://doi.org/10.1121/1.4927411
Xie,X.,&Myers,E.B.(2017).Learningatalkerorlearninganaccent:Acousticsimilarity
constrainsgeneralizationofforeignaccentadaptationtonewtalkers.Journalof
MemoryandLanguage,97,30–46.https://doi.org/10.1016/j.jml.2017.07.005