Fossil Markers of Language Development Phonological_deafness.in_royaumont_OUP

Fossil markers of language development:phonological deafnesses in adult speech processing*

Emmanuel Dupoux and Sharon Peperkamp

* October, 1999. Reprinted by the authors from B. Laks and J. Durand (Eds) (to appear). Cognitive Phonology. Oxford

University Press. The names of the authors are in alphabetic order. Research for this paper was supported by a Fyssengrant to the second author. We would like to thank Leticia Maria Sicuro Correa, Anne Christophe, Franois Dell, JacquesMehler, and Franck Ramus for comments and discussion. Laboratoire de Sciences Cognitives et Psycholinguistique, Ecole des Hautes Etudes en Sciences Sociales, 54 Bd

Raspail, 75006 Paris, France. Email: [email protected]. Web: http://www.ehess.fr/centres/lscp/persons/dupoux Dpartement des Sciences du Langage, Universit Paris 8, 2 Rue de la Liberte, 93535 Saint Denis. Email:

[email protected]

Abstract

The sound pattern of the language(s) we have heard as infants affects the way in which we perceivelinguistic sounds as adults. Typically, some foreign sounds are very difficult to perceive accurately,even after extensive training. For instance, native speakers of French have troubles distinguishingforeign words that differ only in the position of main stress, French being a language in which stress isnot contrastive.

In this paper, we propose to explore the perception of foreign sounds cross-linguistically in orderto understand the processes that govern early language acquisition. Specifically, we propose to test thehypothesis that early language acquisition begins by using only regularities that infants can observe inthe surface speech stream (Bottom-Up Bootstrapping), and compare it with the hypothesis that theyuse all possible sources of information, including, for instance, word boundaries (InteractiveBootstrapping).

We set up a research paradigm using the stress system, since it allows to test the various optionsat hand within a single test procedure. We distinguish four types of regular stress systems theacquisition of which requires different sources of information. We show that the two hypotheses makecontrastive predictions as to the pattern of stress perception of adults in these four types of languages.We conclude that cross-linguistic research of adults speech perception, when coupled with detailedlinguistic analysis, can be brought to bear on important issues of language acquisition.

Index: phonological deafness, language acquisition, phonological bootstrapping,stress perception, cross-linguistic studies

1. Introduction

Recent research has demonstrated that languageacquisition begins at a very early age and proceedsat an amazingly fast pace. During the first year oflife, infants find out important parts of thephonological structure of their maternal languageand begin to recognize and store frequent words.One crucial question for models of languageacquisition concerns the mechanisms underlyingthese developments. In particular, the questionarises as to what type of learning algorithm infantsuse to extract phonological and lexicalinformation. The more specific question we are

interested in is whether learning the words ofones language is independent from learning thephonology, or whether the two processes areinterleaved and interdependent.

In this paper, we claim that cross-linguisticstudies of the end-state in adults can shed light onthese developmental questions. We rest our claimon the finding that early exposure to a languagehas a lasting impact on speech processing routinesin adults. That is, listeners use a processingapparatus specifically tuned to their maternallanguage. Consequently, they have a lot ofdifficulty in dealing with sound structures that arealien to the language they heard as infants. They

DUPOUX AND PEPERKAMP 2

display what we call phonological deafnesses;that is, they have troubles discriminatingphonological contrasts that are not used in theirnative language. Moreover, the phonologicaldeafnesses are robust, in that - analogously topatterns of foreign accent in production - they areresistant to learning a second language, and evento specific training. We hypothesize thatphonological deafnesses originate in theacquisition during the first few years of life. Wetherefore propose to look at perception in adultscross-linguistically, in order to gain insight into theacquisition processes that they went throughduring these first years. Crucially, we propose totest the predicted deafnesses in variouslanguages according to different theoreticaloptions regarding early language acquisition.

The outline of this paper is as follows. Insection 2, we expose some background dataregarding, on the one hand, phonologicalprocessing and acquisition, and, on the other hand,lexical processing and acquisition. In section 3, wepresent two hypotheses concerning the relationshipbetween phonological and lexical acquisition, i.e.Bottom-up Bootstrapping and InteractiveBootstrapping. In section 4, we propose to testthese hypotheses by means of a cross-linguisticstudy on the perception of stress by speakers oflanguages that differ crucially in their stress rules.That is, we consider four types of regular stresssystems the acquisition of which requires differentsources of information, and we discuss thepredictions following from the two hypothesesregarding the pattern of stress perception of adults.Finally, a summary and some concluding remarksare presented in section 5.

2. Experimental data

2.1 Phonological processing andacquisition

During the first year of life, infants acquire manyphonological properties of their native languageand lose their sensitivity to phonological contraststhat are not pertinent. In this section, we reviewdata concerning language-specific phonologicalprocessing in adults and its relation tophonological acquisition. In particular, we dealwith the segmental inventory, phonotactics, andsuprasegmentals, respectively.

2.1.1 SegmentsIt has long been known that speech perception isinfluenced by phonological properties of thematernal language (Sapir 1921; Polivanov 1974).Much experimental evidence has been gatheredconcerning this influence. For instance, Goto(1971) has documented that Japanese listeners mapAmerican [l] and [r] onto their own, single, [R]category, and, as a result, have a lot of difficultiesin discriminating between them. Similarly, thecontrast between the retroflex and dental stops [ ] [t

] is very difficult for the English, but not forthe Hindi speaker (Werker & Tees 1984b). Thiscontrast is, in fact, phonemic in Hindi, whereasneither of the stop consonants involved occurs inEnglish; rather, English uses the alveolar stop [t].Note, however, that not all foreign contrasts aredifficult (Polka 1991; Best, McRoberts, & Sithole1988). For instance, Best et al. found that Englishsubjects do not have difficulties in discriminatingdifferent Zulu clicks from one another.

The Perceptual Assimilation Model (Best1994) aims at explaining these differences. Inparticular, it states that a foreign sound isassimilated to an existing segment in the nativelanguage if the acoustic characteristics of bothsounds are close enough. In this case, listeners areonly able to judge whether the foreign sound is agood or a bad exemplar of the segment to which ithas assimilated, but they do not have access to itsdetailed phonetic characteristics. Consequently,two distinct foreign sounds that are assimilated tothe same segment in the native language and thatare equally bad exemplars of this segment will bevery difficult to discriminate. Such is the case forJapanese speakers with the American [l] - [r]contrast, both [l] and [r] being assimilated toJapanese [R], as well as for the English speakerswith the Hindi [ ] [t

] contrast, both [ ] and [t

]being assimilated to English [t]. By contrast,foreign sounds that are too distant from all of thenative segments will not be assimilated at all andlisteners will be able to make fine phoneticdiscriminations between them. This is illustratedby the successful discrimination of different Zuluclicks by native speakers of English.

The effects of the maternal inventory onadult perception are very robust, and resist to theacquisition of a second language. Thus, even earlyand highly fluent bilinguals have difficulties withnon-native vowel contrasts, as shown in Pallier,

FOSSIL MARKERS OF LANGUAGE DEVELOPMENT 3

Bosch & Sebastin-Galls (1997). In this study,subjects are Spanish-Catalan bilinguals who havebegun to acquire their second language betweenfour and six years of age, and have used itextensively thereafter. The two languages differ asto the number of vowel phonemes; Spanish has afive-vowel system, while Catalan uses two morevowels. Crucially, subjects whose native languageis Spanish are shown to use exclusively the morerestricted Spanish vowel system for the purposesof speech perception and spoken word recognition.

Finally, electrophysiological studiessuggest that effects of the phoneme inventoryoccur very early during on-line speech processing.Thus, Ntnen et al. (1997) presented repeatedsequences of the same vowel, followedoccasionally by a change in vowel. They foundthat the Mismatch Negativity (anelectrophysiological component that correlateswith detection of a change in stimulation) ismodulated as a function of whether the vowel ispart or not of the phoneme inventory of thelanguage. That is, 100 to 240 ms. after the vowelonset, the perceptual system is already influencedby the vowel inventory of the native language.

All in all, these data suggest that listenersuse a set of language-specific phoneme categoriesduring speech perception. This raises the questionas to how and when infants acquire their nativesegmental inventory. Concerning the age at whichinfants tune to the inventory of their language,Kuhl et al. (1992) have shown that American andSwedish 6-month-old infants react specifically tovowels that are prototypic in their maternallanguage. Furthermore, Polka & Werker (1994)found that at the same age, infants begin to losetheir sensitivity for non-native vowel contrasts.That is, 6-month-old English-acquiring infants failto discriminate between the German lax vowels /

/and / / as well as between their tense counterparts/y/ and /u/; the first, front, vowel in each pair is notpart of the English inventory. Between 10 and 12months, infants similarly lose the ability todiscriminate non-native consonantal contrasts(Werker & Tees 1984a). Regarding the way inwhich infants acquire their segmental inventory,Kuhl et al. (1997) found that mothers addressingtheir infants produce acoustically more extremevowels than they do when addressing adults,resulting in a stretching of the vowel space. Thisshows that language input to infants provides well-specified information about the phonemes, and

suggests that a purely statistical extractionalgorithm could establish prototypes for the soundsof the native language.

2.1.2 PhonotacticsLanguage phonologies differ in properties otherthan the inventory of phonemes. Notably, theydiffer in the way in which phonemes can co-occurin words. These phonotactic properties appear toinfluence speech processing routines, in that non-words with a high phonotactic probability areprocessed faster and more accurately than non-words with a low phonotactic probability (Brown& Hildum 1956; Vitevitch et al. 1997; Gathercoleet al. 1999). Furthermore, phonotactics have beenshown to bias the perception of individualsegments. For instance, Massaro & Cohen (1983)found that synthetic stimuli that are ambiguousbetween /r/ and /l/ tend to be perceived as /r/ whenpreceded by /t/ and as /l/ when preceded by /s/.The interpretation given by Massaro and Cohen isthat perception is biased towards segments thatyield the legal clusters /tr/ and /sl/, rather than theillegal clusters /tl/ and /sr/ (see also Pitt 1998).Similarily, Hall et al. (1998) found that illegalonset clusters in French are perceived as legalones. In particular, illegal /dl/ and /tl/ areperceived as legal /gl/ and /kl/, respectively.

It has even been reported that in somecontexts, illegal sequences of phonemes yieldperception of illusory segments. For instance,Japanese syllables cannot have complex onsets(except for consonant-glide onsets) and cannothave codas (except for nasal consonants and thefirst half of geminates). Dupoux et al. (1999)found that Japanese subjects report the presence ofan epenthetic vowel [u] between consonants innon-words like [ebzo]. They also found thatJapanese subjects have problems discriminatingbetween, for instance, [ebzo] and [ebuzo]. Thiswas found even in subjects who were quiteproficient in French, a language which authorizesboth coda consonants and complex onsets.Moreover, in an electrophysiological study,Dehaene-Lambertz, Dupoux & Gout (in press)found that the effect for phonotactics arises asearly as that of the phoneme inventory investigatedby Ntnen et al. (1997) and reported above.These results, then, suggest that phonotactics playa role so important as to create the illusoryperception of segments.


As to infants sensitivity to phonotacticproperties, there is evidence that it equally arisesduring the first year of life. For instance, Friederici& Wessels (1993) showed that 9-month-old Dutchinfants prefer to listen to phonotactically legalwords rather than to illegal ones. Similarly,Jusczyk, Luce & Charles-Luce (1994) found that9-month-old American infants, when listening tomonosyllabic non-words, prefer those with a high-probability phonotactic pattern rather than thosewith a low-probability phonotactic pattern.Jusczyk et al. (1993) reported, furthermore, that 9-month-old American infants listen longer tounfamiliar English words than to Dutch words.The latter contain segments and sequences that areillegal in English, suggesting again that infants ofthis age are sensitive to the phonotactics of theirlanguage. This is corroborated by the finding thatno differences are found when the stimuli are low-pass filtered, hence did not contain any segmentalinformation. Note, however, that we do not yetknow whether at the same age, the presentation ofillegal clusters yields the perception of illusorysegments as documented for adults.

2.1.3 SuprasegmentalsFinally, languages differ in suprasegmentals and inparticular, in the use of word-internal prosody.Two examples might illustrate this. First, considerstress in Spanish and French. In Spanish, stressfalls on one of the words last three syllables(Navarro Tomas 1965), and there are minimalpairs of words that differ only as far as the locationof stress is concerned, for instance bbe (s/he)drinks - beb baby. In French, by contrast, stressdoes not carry lexical information. Rather, itpredictably falls on the words final vowel (Schane1968; Dell 1973). Thus, speakers of Spanish haveto process and represent stress to identify thelexical item(s) intended by the speaker. Speakersof French, by contrast, do not need to processstress, at least not in the same way.1 Dupoux et al.(1997) found that French subjects are deaf tostress. That is, French subjects as opposed toSpanish subjects - exhibit great difficulties indiscriminating non-words that differ only in thelocation of stress. Another example of word-internal prosody concerns the use of vowel length.

1 Rather, stress may be used as a cue to word

segmentation (Trubetzkoj 1939; Rietveld 1980).

In Japanese, but not in French, vowel length iscontrastive. Thus, we find minimal pairs inJapanese such as [to] door and [too] tower. InFrench, by contrast, vowel length is not used tomake lexical distinctions. Accordingly, Dupoux etal. (1999) found that French, but not Japaneselisteners have great difficulties in distinguishingbetween non-words that differ only in vowellength.

Little is known about the acquisition ofsuprasegmentals in young infants. However, thereis evidence that newborns are already sensitive tosome global suprasegmental properties. Inparticular, it appears that on the basis of theseproperties, they distinguish between their nativelanguage and a foreign language, as well asbetween different foreign languages. Thus, Mehleret al. (1988, 1996) found that French infantsdiscriminate both between French and Russianutterances and between English and Italianutterances. The stimuli were low-pass filtered,indicating that discrimination can be achieved onthe basis of suprasegmental information only.Similarly, Moon, Cooper & Fifer (1993) foundthat English-acquiring newborns prefer to listen toEnglish rather than to Spanish sentences, whileSpanish-acquiring newborns show the reversepreference pattern (see also Nazzi, Bertoncini &Mehler 1998).

Furthermore, Jusczyk et al. (1993) showedthat 6-month-old American infants prefer to listento English rather than to Norwegian words, whilethey fail to show a preference for English words asopposed to Dutch words. These results also holdwhen the stimuli are low-pass filtered, suggestingthat infants are sensitive to suprasegmentalproperties that are typical of their native language;English and Dutch indeed share many of theseproperties, whereas Norwegian suprasegmentalsdiffer substantially from those of English. Alongthe same lines, Jusczyk, Cutler, & Redantz (1993)found that 9-month-old American infants prefer tolisten to disyllables with the metrical pattern whichis predominant in English, i.e. with stress on thefirst rather than on the second syllable. However,this experiment has not been carried out with alanguage that shows the reverse metrical pattern. Itcould, therefore, be the case that the obtainedpreference of American infants stems from auniversal bias, rather than being related to thepredominant metrical pattern of disyllables in thenative language.


Finally, it is currently unknown at whatage infants begin to exhibit the type of deafnessto foreign prosodic contrasts that has been foundin adults.

2.2 Lexical processing and acquisition

During the first year of life, infants not onlyacquire many phonological properties of theirlanguage, they also begin to build a lexicon. In thissection, we review data regarding lexicalprocessing and acquisition.

It has been argued that in adult speechprocessing, function words and content words areprocessed differently. For instance, Friederici(1985) reported that in a word-monitoringexperiment, responses to function words werefaster than responses to content words. Thissuggests that function words and content words arenot stored together. In fact, given that the set offunction words is extremely limited, searchprocedures within this set are faster than thatwithin the set of content words. Similarly, Neville,Mills & Lawson (1992) found that the brain elicitsqualitatively different electrophysiologicalresponses to function words and content words,respectively.

There is recent experimental evidence thatthe distinction between function words and contentwords is acquired early in life. In particular,Shady, Jusczyk & Gerken (1988) (cf. Jusczyk, inpress) found that 10-month-old American infantslisten shorter to passages in which function wordsare replaced by non-words having the samephonological properties. By contrast, they do notlisten shorter if content words are replaced by non-words having the same phonological properties.This suggests that at this age, infants not onlymake a distinction between function words andcontent words, but also recognize the actualfunction words of English. By contrast, they do notknow the semantics of these words, as evidencedby a follow-up experiment. That is, infants listenedequally long to normal passages and to passages inwhich the function words were exchanged amongeach other, leading to ungrammatical sentences.

Gerken (1996) and Morgan, Shi &Allopenna (1996) proposed that infants acquirefunction words before content words and that theydo so on the basis of phonological cues.Specifically, function words typically sharephonological properties that set them apart from

content words. In English, for instance, functionwords are characterized by having a short durationand low relative amplitude, a simple syllablestructure, and centralized vowels. Moreover, theytend to occur utterance-initially. Shi (1995)showed that taken together, these cues aresufficient for a self-organizing neural network toclassify words as function words or content wordswith an accuracy of 85-90 percent. Shi, Morgan &Allopenna (1998) obtained similar results with twounrelated languages, i.e. Mandarin Chinese andTurkish, suggesting that function words canuniversally be set apart from content words on thebasis of acoustic, phonological and distributionalcues only.

As to the beginning of the compilation of alexicon of content words, Jusczyk & Aslin (1995)showed that it could lie as early as 7 months ofage. That is, infants of this age listen longer topassages containing a word to which they arehabituated than to passages that do not containsuch a word. The same results are obtained ifinfants are habituated to passages containingseveral instances of certain words and tested onthese words in isolation. Thus, infants listen longerto words that are contained in the passages theyheard previously than to words that are notcontained in the passages. Moreover, wordsappeared to be stored in a detailed phoneticrepresentation. For instance, when trained on cup,infants show no recognition of tup, which differsonly as far as the place of articulation of the firstsegment is concerned.

Benedict (1979) reported - on the basis ofcomprehension tests as well as observational datain mothers diary notes - that English-learninginfants of 10 months comprehend around 10words; this figure grows to around 40 at 12 month,and to 100 or more at 16 months. Recentexperimental work is consistent with this report.For instance, Hall & Boysson-Bardies (1994)found that at 10 months of age, French infantsprefer to listen to a list of 12 familiar rather than toa list of 12 unfamiliar words. More surprisingly,Mandel, Jusczyk & Pisoni (1995) reported that4-month-old infants recognize their own name.Specifically, infants were shown to prefer to listento their own name rather than to names with thesame number of syllables and the same stresspattern. Moreover, such preference disappears ifthe initial phoneme of the infants name is changed(Nazzi & Jusczyk, personal communication).


Hence, recognition of the proper name of theinfant seems to be based on precise segmentalinformation rather than on global prosodicproperties.

2.3 Open questions

As it is apparent in the above review, bothphonological acquisition and lexical acquisitionbegin very early in life and develop rapidly.However, many questions regarding themechanisms that are responsible for such speedyacquisition remain open and the proposed learningmechanisms contain paradoxical circularities. Onthe one hand, the acquisition of certainphonological properties, such as phonotactics orthe typical prosodic shape of words, seems torequire the prior acquisition of a lexicon of areasonable size in order to extract some stablestatistics. On the other hand, lexical acquisitionitself seems to require some prior phonologicalknowledge, such as phoneme categories andlanguage-specific word boundary cues. This raisesquestions regarding the relative time course ofphonological and lexical acquisition, as well as thepotential interactions between the two processes.In the next section, we examine the varioustheoretical alternatives.

3. Setting up the research framework

3.1 Assumptions and hypotheses

Before discussing different theoretical pathways inearly language acquisition, let us first describethree underlying assumptions (cf. Mehler, Dupoux& Segui 1990). First, infants have an innateuniversal phonetic representation. Thisrepresentation encodes speech sounds and keepsall phonetic distinctions that can in principle beused contrastively, while reducing the importanceof non-linguistic variables such as talker voice,length of vocal tract, and background noise.Second, infants similarly acquire a lexicon of wordforms during the first years of life. This lexicondistinguishes between content words and functionwords. Third, during the first years of life, infantsacquire a language-specific prelexicalrepresentation. This representation is intermediatebetween the universal phonetic representation andthe lexicon. It is discrete and encodes only a smallsubset of those segmental and suprasegmental

distinctions that are available at the universalphonetic level. It thus specifies the format underwhich the lexical items are stored. Once acquired,the prelexical representation is assimilatory; thatis, foreign sound patterns are assimilated to theclosest native sound pattern.2 This holds not onlyfor the segmental inventory, as in the PerceptualAssimilation Model of Best (1994), but also for allother aspects of phonological structure, such asphonotactics and suprasegmentals. For instance,foreign words that contain illegal syllabic structurewill be regularized through the assimilation toexisting syllables in the native language (Dupouxet al. 1999). Finally, the prelexical representationis crystallized in the adult; that is, it remains stableeven after extensive exposure to a secondlanguage. All these properties result in patterns ofphonological deafness for certain non-nativecontrasts, i.e. those contrasts involving sounds thatare assimilated to a single sound pattern in theprelexical representation.

Within this framework, the acquisitionproblem can be stated as follows: At birth, thelanguage-specific prelexical representation andlexicons of content words and function words areunavailable to the infant, while a few years later,the acquisition of these components has beencompleted; how, then, has this been accomplished?The paradox is that each one of these componentsof the processing system seems to require the prioracquisition of the other one before it can itself beacquired. On the one hand, lexical acquisitionseems to require a language-specific prelexicalrepresentation. In fact, a given word can surface ina near infinity of phonetic forms that - if thelexicon were constructed on the basis of auniversal phonetic representation - would all bemapped onto separate lexical entries. On the otherhand, in order to acquire a prelexicalrepresentation, infants seem to need minimal pairsin order to decide which variation is pertinent; thatis, they need to have a lexicon. This bootstrappingproblem is one of the most puzzling questions inearly language development. In the following, wediscuss two theoretical possibilities to solve thepuzzle. The first one we dub Bottom-Up

2 If the foreign sound is too far away from any native

sound (as in the case of Zulu clicks for Englishspeakers), it is not assimilated but perceived on a purelyphonetic basis (see Best et al. 1988).


Level III

universal phoneticrepresentation

Level IV

Level I

Level II ! " # $ % & ' ( ) * +language-specific

prelexicalrepresentation

lexicon

function words content words

Figure 1: compilation of prelexical representation in earlylanguage acquisition

Bootstrapping, the second one InteractiveBootstrapping.

In Bottom-Up Bootstrapping, acquisitionbegins on the basis of the acoustic signal only, andproceeds sequentially; once a component isacquired, the next component comes into play,using only information available at lower levels(cf. Mehler, Dupoux & Segui 1990; Christophe &Dupoux 1996; Mehler et al. 1996). In ourframework, this would mean that the prelexicalrepresentation would be acquired first, exclusivelyon the basis of information that is available in theuniversal phonetic representation. Thisrepresentation would then be used to extractwords. Given that only partial linguisticinformation would be available for the compilationof the prelexical representation, it would notnecessarily be fully optimized for the nativelanguage. In particular, in the absence of lexicalinformation, evidence based on the presence orabsence of minimal pairs would not be taken intoaccount. Allophones, then, would generally beencoded separately.3 In other words, the prelexicalrepresentation would contain some redundancy.

In Interactive Bootstrapping, by contrast,lexical and phonological acquisition beginsimultaneously and interact with one another. Inour framework, this would mean that part of thephonological representation would be acquiredbottom-up through inspection of the phoneticrepresentation, and part would be acquired on thebasis of lexical information. Similarly, the first

3 In Peperkamp & Dupoux (submitted), we argue that

some allophones can be detected on the basis of thephonetic representation only.

words would be acquired from the universalphonetic representation, while they would berecoded more and more abstractly as morephonology becomes available. Thus, in InteractiveBootstrapping, each component could in principlebe optimized for all language-specific properties,all relevant sources of information being available.As to the prelexical representation, its acquisitionwould continue until all redundancy would havebeen removed. In other words, all and only thosedistinctions that are used contrastively would beencoded.

3.2 The acquisition of the prelexicalrepresentation: a typology

In the remaining part of this paper, we compareBottom-up Bootstrapping and InteractiveBootstrapping by focussing on the acquisition ofthe prelexical representation. The question weinvestigate concerns the types of information thatallow infants to compile the prelexicalrepresentation and decide whether a givenphonological distinction is to be kept or not in thisrepresentation. The prelexical representation isshown in Figure 1; the square brackets representthe utterance boundaries. The four types ofinformation that could be taken into accountduring the compilation of the prelexicalrepresentation are numbered in increasing order ofcomplexity. First of all, Type I information can beextracted from the universal phoneticrepresentation. Type II information can beextracted from the prelexical representation itselfby making reference to already acquiredphonological properties. Type III information


regards the distribution and shape of functionwords. Finally, Type IV information concerns thelocation of content word boundaries in theutterance.

Bottom-up Bootstrapping and InteractiveBootstrapping make different predictions as to thetypes of information that can affect thecompilation of the prelexical representation. Thatis, according to Bottom-up Bootstrapping, onlylow-level information, i.e. Type I and Type II, canaffect prelexical acquisition; according toInteractive Bootstrapping, by contrast, high-levelinformation, i.e. Type III and IV, might also comeinto play. Deafnesses originating from Type Iand Type II information, then, would support bothBottom-up Bootstrapping and InteractiveBootstrapping, while deafnesses originating fromType III and Type IV information would provideevidence in favor of Interactive Bootstrappingonly.

We can now classify phonologicalgeneralizations according to the type ofinformation that is required in order for them to beobserved. Thus, a phonological generalization ofType 1 is a generalization that can be observed onthe basis of Type 1 information, and so forth. Weexemplify one by one the four types ofgeneralizations and determine the correspondingdeafnesses that are or might be attested. First,regarding Type I, certain generalizations relevantto the prelexical representation are directlyobservable on the basis of the universal phoneticrepresentation. For instance, there is someevidence that the distribution of vowels aroundlanguage-specific prototypes is observable at anacoustic level (Kuhl et al. 1997). Consequently,one observes an early prelexical acquisition ofprototypic vowels in the language (Kuhl et al.1992), yielding language-specific deafnesses tocontrasts of vowels that are phonetically differentbut map onto the same prototype (e.g. Pallier,Bosch & Sebastin-Galls 1997).

Second, Type II generalizations require theprior acquisition of other aspects of the phonologyof the language. For instance, German has bothvoiced and unvoiced obstruents, but syllable-finally only the latter can occur (Wiese 1996). Inorder for German infants to observe this regularity,they should have access to German syllablestructure. Thus, once syllable structure is encodedin the prelexical representation, the regularityconcerning the absence of syllable-final voiced

obstruents can be made. Provided that the latterobservation feeds back on the compilation of theprelexical representation, the voicing feature willnot be encoded in the prelexical representation forsyllable-final consonants. German adults, then, arepredicted to exhibit a deafness to syllable-finalvoicing contrasts.

Similarly, in Italian, vowel length isallophonic: vowels are lengthened if and only ifthey occur in an open stressed syllable (Vogel1982). In order to find this regularity, infants needto have acquired both the distinction between openand closed syllables and the distinction betweenstressed and unstressed syllables in their language.Once such information is contained in theprelexical representation, the higher-orderobservation (that is the correlation of syllablestructure, stress, and vowel length) becomesavailable. Again, provided that this higher orderobservation can feed back on the compilation ofthe prelexical representation, vowel length ispredicted to be removed from the prelexicalrepresentation. As a consequence, Italian-speakingadults should have difficulties distinguishingbetween short and long vowels; that is, they shouldexhibit a deafness to vowel length contrasts.

Third, Type III generalizations can beobserved only once the distinction between contentwords and function words is made. For instance, inDutch, words can begin with any vowel exceptschwa. The prohibition on word-initial schwa,though, does not hold for function words, witnessexamples such as een [

,

n] a and het [,

t] it(Booij 1995). Once infants can extract functionwords out of the speech stream, they can observethat the remaining strings do not begin with schwa.If this regularity is taken into account during thecompilation of the prelexical representation, wemight expect that in adult speech perception,foreign words with an initial schwa aremisperceived and assimilated to a legal word-initial vowel.

Fourth, in order for Type IVgeneralizations to be observed, the boundaries ofcontent word have to be available. For instance, inNorthern varieties of Standard Italian, the contrastbetween [s] and [z] is allophonic; that is, withinwords, [z] surfaces before voiced consonants andin intervocalic position, while [s] occurseverywhere else (Camilli 1965; Nespor & Vogel1986). In intervocalic position within phrases,then, both [z] and [s] occur, depending on the


presence or absence of a word boundary within thesequence. This is illustrated in (1).

(1) a. bu[z]ecca bovine tripesb. bu[s] ecologic ecological bus

Therefore, infants need to segment utterances intoseparate words in order to find the generalizationconcerning the distribution of [s] and [z].4 If wordboundaries or lexical knowledge can influence thecompilation of the prelexical representation, onemight expect that intervocalic [z] will be recodedas underlying /s/. Italian-speaking adults, then,should exhibit deafness to the contrast betweenintervocalic [s] and [z].

The inventory presented above does notexhaust the type of deafnesses that couldtheoretically arise. We might envision cases whereonly knowledge of syntactic categories ormorphological decomposition can inform theprelexical level that some distinction is irrelevantand hence can be removed from the representation.The predictions are the same: if such high-levelinformation is available to infants while they arecompiling the prelexical phonologicalrepresentation, then a deafness to this type ofdistinction should be observed in adults.

In the examples given above, differentphenomena (neutralization, allophonic variation,phonotactics) were involved. Hence, testing thesefour types of predicted deafness would involveusing different materials and experimentalparadigms. In the remaining part of this paper, weexemplify the four-way distinction with the stresssystem, which allows to construct a singleexperimental test that can be applied cross-linguistically.

4. A cross-linguistic test case: stressdeafness

One domain in which we find limited capacities toperceive phonological contrasts that are notpertinent in the native language is that of stress.For instance, recall from section 2.1.3 that Frenchlisteners have great difficulties perceiving thedifference between non-words that aredistinguished only with regard to the position of

stress, such as vsuma - vasma - vasum(Dupoux et al. 1997). In contrast, Spanish subjectshave no problem in making this distinction.

In our framework, the stress deafness inFrench adults could arise as a consequence of theway infants acquire stress. In French, stress ispredictable in that main stress always falls on thewords final syllable (Schane 1968; Dell 1973), ageneralization that can be inferred by infants onthe basis of the universal phonetic representation.Indeed, since as we will show in detail in section4.1 all utterances end with a stressed syllable,infants can deduce that stress is not contrastive bypaying attention to the ends of utterances only.Therefore, they will not encode stress in theprelexical representation, and as adults they willexhibit a deafness to stress contrasts. In Spanish,by contrast, stress falls on one of the words lastthree syllables. Although there are certainrestrictions on stress placement, the existence ofminimal pairs such as beb baby bbe drinksshows that there is no way to reliably predict thestress location in all cases. Consequently, stresswill be encoded in the prelexical representation.

French and Spanish hence represent twoextreme cases: one in which stress is Type Ipredictable, and one in which it is unpredictable.In this section, we will explore languages thatdisplay stress regularities of the four types definedin section 3. The patterns of stress perception byadult speakers of these languages will help todetermine the types of information that are takeninto account in the compilation of the prelexicalrepresentation during infancy, and hence to drawconclusions regarding the validity of the twobootstrapping hypotheses.

Before going into this, however, we shoulddefine what is meant exactly by studying theperception of stress cross-linguistically. We willexamine stress as it is realized in the subjectsnative language, and hence test the perception ofthe acoustic cues that are typical of stress in thelanguage under consideration. Crucially, we makesure that we do not manipulate variables that couldbe perceived as something else than stress in thatlanguage. The acoustic correlates of stress areloudness, pitch and duration (Lehiste 1970), butnot every language uses all three cues to realize

4 We abstract away from problems posed by prefixes for

the distribution of [s] and [z] (see, e.g., Nespor & Vogel1986; Peperkamp 1997)


stress. For instance, in languages with contrastivevowel length, duration is avoided as a correlate ofstress (Hayes 1995). Therefore, when processing aforeign language with duration as phoneticcorrelate of stress, speakers of languages withcontrastive length might map stressed vowels ontolong vowels and unstressed vowels onto shortvowels. Thus, they can assimilate stress to length,and consequently, stress deafness will not beobserved. Similarly, tone languages typically avoidpitch as a correlate of stress. Thus, whenprocessing a foreign language with pitch asphonetic correlate of stress, speakers of tonelanguages might map stressed vowels onto hightone vowels and unstressed vowels onto low tonevowels. In other words, they can assimilate stressto tone, and again, stress deafness will not beobserved. For conveniences sake, we willcontinue to use the term stress deafness inreferring to deafness to the phonetic correlate(s)of stress as present in the subjects nativelanguage. For an experimental paradigm thatenables testing the perception of stress in adultspeakers, see Peperkamp, Dupoux & Sebastin-Galls (1999).

We will now turn to languages withdifferent stress rules and spell out our predictionsconcerning the perception of stress by nativespeakers of these languages. The languages areclassified according to the type of regularitypresented by their stress rules.

4.1 Type I

First, let us look at languages in which the stressrule corresponds to a Type I generalization. Frenchis a case at hand, the stress rule (stress the finalvowel) making reference only to phonetic notions.Notice that even clitics, which are typicallyunstressed elements, attract stress if they arephrase-final.5 This is illustrated in (2).

(2)a. coupez [kup] cutIMP-PLb. coupez-les [kupel] cutIMP-PL-themc. coupez-vous-en [kupevuza

-

. ] cutIMP-PL-yourselfPL-DAT-of them

5 Exceptions are je I and ce it. These clitics have

schwa as their vowel, which is deleted in phrase-finalposition, as in suis-je [s/

0

] am I and est-ce [1 2

s] is it.In other words, these two clitics do not undermine thesurface observability of stress falling on the final voweleither.

Moreover, although eurhythmic principles inducedestressing rules in French, the final and strongeststress of an utterance is never reduced (Dell 1984).Hence, neither the occurrence of enclitics nordestressing interferes with the phoneticobservability of the stress rule. Therefore, allutterances in French have stress on their finalsyllable.6 Infants, then, can infer that stress is notcontrastive, hence, needs not be encoded in theprelexical representation.

Pitta-Pitta (Blake 1979; Hayes 1995) isanother example of a language with a stress rulethat should give rise to Type I deafness. ThisAustralian language has fixed word-initial stress(3), and there are no proclitics.

(3)a. krru boyb. mlilpu eyebrowc. ypariri young man

Pitta-Pitta has no monosyllabic words; stress clash,therefore, never occurs. Thus, all utterances havestress on the first syllable. Given that stress ispredictable and observable on the basis of theuniversal phonetic representation, both theBottom-up Bootstrapping hypothesis and theInteractive Bootstrapping hypothesis predict thatstress is not encoded in the prelexicalrepresentation and that speakers of Pitta-Pittaexhibit stress deafness.

4.2 Type II

An example of a stress system involving a Type IIregularity is presented by Fijian (Schtz 1985;Dixon 1988; Hayes 1995). In this Austronesianlanguage, word stress falls on the final syllable if itis heavy; otherwise stress is penultimate. Thelanguage has only two syllable types, (C)VV and(C)V, where the former is heavy and the latter is

6 In southern varieties of French, words can end in an

unstressed schwa. In these varieties, stress thus falls onthe words last full vowel. The acquisition of the stressregularity, then, is more complex, in that infants firsthave to acquire the difference between full vowels andschwa. Consequently, the stress deafness is of thesecond rather than of the first type.

Note also that, alternatively, French has beencharacterized as having phrasal stress, with stress fallingon phrase-final syllables (Grammont 1965). Thequestion as to whether French has word stress or phrasal


light. Suffixes are within the stress domain,7 andthere are no enclitics. Examples from Dixons(1988) description of the Boumaa dialect are givenin (4); syllable boundaries are indicated by dots.(4) a. l.a lu..ca vomit vomit onTR b. te. 3 e.v 4 te. 5 e.v 4 .na start - startTR c. pu.lu pu.lu.na be covered - coverTR

The language permits monosyllabic wordsprovided they are heavy. If a monosyllable ispreceded by a word ending in a long vowel or adiphthong, a stress clash arises. We do not know ifand how stress clash is resolved in Fijian. Clearly,if the second stress undergoes destressing,utterance-final stress clash configurations disruptthe surface stress pattern. Abstracting away fromthis potential confound, however, we canformulate the following surface generalization: inutterances that end in a word with a final longvowel or a diphthong, the final syllable is stressed,while in utterances that end in a word with a finalshort vowel, the penultimate syllable is stressed.Once infants have acquired the distinction betweenheavy and light syllables, they can observe thestress regularity. Speakers of Fijian, then, arepredicted to exhibit stress deafness if thisregularity is taken into account during thecompilation of the prelexical representation Aswith the Type I regularities, this follows from boththe Bottom-up Bootstrapping hypothesis and theInteractive Bootstrapping hypothesis.

4.3 Type III

In many languages, stress is predictable andobservable modulo the occurrence of clitics. Infact, contrary to the situation in French describedin section 4.1, clitics are typically unstressed,regardless of their position in the utterance.Consider, for instance, the case of Hungarian. Inthis language, stress falls on the word-initialsyllable, and clitics are systematically unstressed(Vago 1980). This is illustrated in (5).

(5) a. emberek [mberek] menb. az emberek [azmberek] the men

stress is irrelevant to our point, given that under bothassumptions, all utterances end in a stressed syllable.7 Some suffixes and sequences of suffixes, however,

form separate stress domains. That is, they behave asindependent words phonologically (cf. Dixon 1988).

In Hungarian, then, utterances that begin with aclitic have stress on the second syllable, while allother utterances have stress on the first syllable.Hungarian has monosyllabic content words, butstress clash resolution does not interfere with thissurface stress pattern (Vogel 1988).

In order for prelexical infants to discoverthe stress rule of Hungarian, they should strip offutterance-initial function words and look forgeneralizations in the remaining string. That is,after removing the initial function word(s), infantscan discover that the remaining string of contentwords always begins with a stressed syllable.Bottom-up Bootstrapping states that such adiscovery cannot affect the compilation of theprelexical representation. In fact, according to thishypothesis, only low-level phonetic andphonological information is taken into accountduring the compilation of the prelexicalrepresentation. As a consequence, adult speakersof Hungarian should not exhibit stress deafness,even though stress is not contrastive in theirlanguage. By contrast, Interactive Bootstrappingstates that all possible sources of information including those pertaining to the distinctionbetween function words and content words - isused in order to determine which contrasts arepertinent in the language. According to this latterhypothesis, stress would not be encoded in theprelexical representation; hence, adult Hungariansshould exhibit stress deafness.

4.4 Type IV

Certain stress rules are observable only if theboundaries of content words are available.Consider, for instance, Piro, an Arawakanlanguage in which word stress is on thepenultimate syllable (Matteson 1965; Hayes 1995).Examples from Matteson (1965) are given in (6).

(6) a. ns genipab. wlo rabbitc. rutxtxa he observes tabood. t6 iyahta he cries

Given that there are no enclitics, the followinggeneralization emerges. In utterances ending in amonosyllabic word, the final syllable is stressed,whereas in all other utterances, the penultimatesyllable is stressed. In order to extract the ruleregarding penultimate stress, prelexical infantsshould have access to content word boundaries.


Therefore, the Interactive Bootstrappinghypothesis but not the Bottom-up Bootstrappinghypothesis predicts that speakers of Piro exhibitstress deafness, since only in the former canword boundaries influence the compilation of theprelexical representation.8

Another example of a purely phonologicalstress rule that is observable only if the boundariesof content words are available is presented byNyawaygi (Dixon 1983; Hayes 1995). In thisAustralian language, stress is assigned at the leftedge of words, as follows. In words beginning witha heavy syllable, stress falls on the first syllable(7a). In words beginning with a light syllable,stress falls on the second syllable if the wordcontains three or five syllables (7b) and on the firstsyllable if the word contains two or four syllables(7c). Notice that coda consonants are weightless inNyawaygi; heavy syllables are syllables containinga long vowel.

(7)a. heavy initial syllable:7 8 9 u fish9 : ba ; i south

b. light initial syllable, uneven nb. of syllables:bulbri quail

c. light initial syllable, even nb. of syllables:< = a manbya < ala water snake

Stress clash does not occur, since all words areminimally disyllabic and no word has final stress.Furthermore, there are no proclitics. Thus, allutterances whose first word begins with a heavysyllable have initial stress, whereas all otherutterances have stress on either the first or thesecond syllable, depending on the number ofsyllables in the first word. In order to extract thisgeneralization, infants not only need to besensitive to syllable weight, but they also need tohave access to the number of syllables in theutterances first word; hence, they must be able tosegment the first word out of the utterances.Therefore, as with speakers of Piro, nativespeakers of Nyawaygi are predicted to exhibitstress deafness according to Interactive

8 For the sake of simplicity, we deliberately chose a

language without enclitics. Of course, a language withpenultimate stress that has enclitics would fall into thesame class of our deafnesses typology, since lexicalsegmentation includes clitic stripping.

Bootstrapping but not according to Bottom-upBootstrapping.

4.5 A caveat

Before closing our typology of stress deafness,we would like to raise the following caveat. Ourreasoning is based on the assumption that thestress pattern at utterance edges allows infants tomake inferences about the general stress rule oftheir language. For instance, we postulate thatfrom the observation that all utterances end with astressed syllable, infants will infer that all wordsend with a stressed syllable. This, of course, maybe unwarranted. Phonological rules can indeedapply at some phrasal edge only. In that case, aregularity appears at some utterance edge whichdoes not hold, however, for individual words. Anexample is provided by the tonal system of Slave.In this language, there are two tones, high and low.At the end of intonational phrases, the distinctionis neutralized; the final word necessarily surfaceswith a low tone (Rice 1987). If infants were toexamine the end of utterances only, they wouldincorrectly conclude that there is no tone in theirlanguage.

Yet, stress seems to be special, in that tothe best of our knowledge - this kind of situationdoes not arise with stress rules. On the contrary, inthe cases in which stress is modified at a phrasaledge, the modification seems to enhance theregularity of the stress pattern instead of obscuringit. For instance, recall from section 4.1 that inFrench, phrase-final enclitics are stressed. Thispattern reinforces the utterance-final regularitycaused by the word-final stress pattern of French.By contrast, we have found no language in which,analogously to tone in Slave, contrastive stress isneutralized at a phrasal edge. It should be notedthat if our model is correct, there cannot be such alanguage. In fact, if it existed, infants would usethe edge regularity to infer that stress ispredictable, and hence they would become stressdeaf; this, then, would have the effect that stresscontrasts would be lost within one generation.

To conclude, the absence of languageswith contrastive stress that is neutralized at somephrasal edge is quite critical for our model. If suchlanguages exist, we have to revise our proposal, byoffering another, more complex, learning principlethat allows infants to acquire the stress rule oftheir language. Note that even in that case, ourhierarchy of languages would still be pertinent to


describe the relative observability of these stressrules from the infants viewpoint. For instance,stress in French (Type I) would still be easier toacquire than stress in Nyawaygi (Type IV).

5. Summary and conclusion

We have shown that cross-linguistic research ofadult speech perception can be brought to bear onissues of language acquisition. In particular, itmight allow to solve some of the most puzzlingissues in early language development, namely, theparadoxical fact that the acquisition of eachprocessing component seems to require the prioracquisition of the other components.

As a first step to solve the acquisitionparadox, we proposed to distinguish twohypotheses, i.e. Bottom-up Bootstrapping andInteractive Bootstrapping. The former states thatacquisition proceeds step by step; specifically, theacquisition of each level exclusively relies uponinformation that is available in the shallower levelsof representation. According to the latterhypothesis, in contrast, all levels begin to beacquired independently and there is interactionbetween levels until a stable state is attained.

We set up our methodology to explore theacquisition of the prelexical representation andfocussed on the stress system, since it allows auniform way of distinguishing the various sourcesof information that could in principle be taken intoaccount in the compilation of the prelexicalrepresentation. We thus proposed to test whetherstress is encoded in this representation by speakersof a variety of languages. In particular, we arguedthat the presence of stress deafness is anindication of the absence of stress in the prelexicalrepresentation. The underlying assumption is thatstress should only be encoded in the prelexicalrepresentation if it is useful to distinguish lexicalitems. The infants problem is to decide, on thebasis of a limited amount of information, whetherstress should be encoded or not.

We distinguished four classes oflanguages, corresponding to four types ofinformation that infants could use in order to makethis decision. That is, in the first class of languages(Type I), the stress rule can be acquired on thebasis of a universal phonetic representation only;

in the second class (Type II), it can be acquiredonce certain language-specific phonologicalinformation has been extracted; in the third and thefourth class (Type II and IV, respectively), it canbe acquired only after function words and contentwords, respectively, can be segmented out of thespeech stream. We then made contrastingpredictions regarding the presence of stressdeafness in adults of these four language classes.The reasoning is as follows. Whenever stressdeafness is exhibited by speakers of a certainlanguage, we take this as evidence that stress is notencoded prelexically, hence, that the stressregularity is not taken into account during thecompilation of the prelexical representation.According to the Bottom-Up Bootstrappinghypothesis, only the universal phonetic and theprelexical representation itself can enter the sceneduring the compilation of the prelexicalrepresentation, hence only deafnesses of Types Iand II are predicted. In contrast, the InteractiveBootstrapping hypothesis states that all processingcomponents are allowed to interact till a stablestate is attained. Hence, one would expect all fourtypes of deafnesses to be observed. A summaryof the languages with their characteristics and thevarious predictions is given in Table 1.

The empirical investigation that wepropose thus allows to test models of earlylanguage acquisition. Such empirical investigationis now under way.

As a final note, we would like to mentionthat although we have focussed on the acquisitionof the prelexical representation, our approach canbe transposed to other processing components thatare involved in early language acquisition, such asword segmentation or lexical access. Provided thatthese components show clear language-specificproperties in adults, a cross-linguistic comparativestudy can probe what types of information are usedby infants to compile such language-specificroutines. In other words, the comparativepsycholinguistics approach using only adult data isa potentially useful tool that can be added to thepanoply of techniques currently used to investigatefirst language acquisition.


Table 1: Summary of predictions for the four classes of stress systems.Regularity Language Stress Rule Interfering

DestressingClitics Stress Deafness

French stress final syllable no stressedenclitics

attested (Dupoux etal. 1997)

Type I

Pitta-Pitta stress initial syllable notapplicable

no

procliticspredicted by both

Bottom-up andInteractive

BootstrappingType II Fijian stress heavy penultimate,

otherwise antepenultimatesyllable

? no enclitics predicted by bothBottom-up and

InteractiveBootstrapping

Type III Hungarian stress first syllable no unstressedproclitics

predicted only byInteractive

BootstrappingPiro stress penultimate syllable ? no enclitics predicted only by

InteractiveBootstrapping

Type IV

Nyawaygi stress initial syllable if heavyor if word contains even

number of syllables;otherwise, stress second

syllable

notapplicable

no enclitics predicted only byInteractive

Bootstrapping

References

Benedict, H. 1979 Early lexical development:comprehension and production. Journal of ChildLanguage 6. 183-200.

Best, C. 1994 The emergence of native-languagephonological influence in infants: A perceptualassimilation model. In J. Goodman and H.Nusbaum, eds. The Development of SpeechPerception: The Transition from Speech Sounds toSpoken Words. 167-224. Cambridge, MA: MITPress.

Best, C., G. McRoberts, and N. Sithole, 1988Examination of perceptual reorganization fornonnative speech contrasts: Zulu clickdiscrimination by English-speaking adults andinfants. Journal of Experimental Psychology:Human Perception and Performance 14. 345-360.

Blake 1979 Pitta-Pitta. In R. Dixon and B. Blake, eds.Handbook of Australian Languages, Vol. 1. 182-242. Amsterdam: John Benjamins.

Booij, G. 1995 The Phonology of Dutch. Oxford:Clarendon Press.

Brown, R. and D. Hildum 1956 Expectancy and theperception of syllables. Language 32. 411-419.

Camilli, A. 1965 Pronuncia e grafia dellitaliano.Firenze: Sansoni.

Christophe, A. and E. Dupoux 1996 Bootstrappinglexical acquisition: the role of prosodic structure.The Linguistic Review 13. 383-412.

Dell, F. 1973 Les Rgles et les Sons. Paris: Hermann.

Dell, F. 1984 Laccentuation des phrases en franais. InF. Dell, D. Hirst and J.-R. Vergnaud, eds. FormeSonore du Language. 65-122. Paris: Hermann.

Dixon, R. 1983 Nyawaygi. In R. Dixon and B. Blake,eds. Handbook of Australian Languages, Vol. 3.430-531. Amsterdam: John Benjamins.

Dixon, R. 1988 A Grammar of Boumaa Fijian. Chicago:The University of Chicago Press.

Dehaene-Lambertz, G., E. Dupoux and A. Gout (inpress) Electrophysiological correlates ofphonological processing: a cross-linguistic study.Journal of Cognitive Neuroscience.

Dupoux, E., K. Kakehi, Y. Hirose, C. Pallier, and J.Mehler 1999 Epenthetic vowels in Japanese: aperceptual illusion? Journal of ExperimentalPsychology: Human Perception and Performance25. 1-11.

Dupoux, E., C. Pallier, N. Sebastian, and J. Mehler 1997A destressing deafness in French? Journal ofMemory and Language 36. 406-421.

Friederici, A. 1985 Levels of processing and vocabularytypes: Evidence from on-line comprehension innormals and agrammatics. Cognition 19. 133-166.

Friederici, A. and J. Wessels 1993 Phonotacticknowledge of word boundaries and its use in infantspeech perception. Perception and Psychophysics54. 287-295.

Gathercole, S., C. Frankish, S. Pickering, and S. Peaker1999 Phonotactic influences on short termmemory. Journal of Experimental Psychology:Learning, Memory and Cognition 25. 84-95.


Gerken 1996 Phonological and distributionalinformation in syntax acquisition. In J. Morganand K. Demuth, eds. Signal to Syntax.Bootstrapping from Speech to Grammar in EarlyAcquisition. 411-425. Mahwah, N.J.: LEA.

Goto, H. 1971 Auditory perception by normal Japaneseadults of the sounds l and r. Neuropsychologia9. 317-323.

Grammont 1965 Trait de phontique. Paris: Delagrave.Hall, P. and B. de Boysson-Bardies 1994 Emergence of

an early receptive lexicon: infants' recognition ofwords. Infant Behavior and Development 17. 119-129.

Hall, P., J. Segui, U. Frauenfelder, and C. Meunier1998 Processing of illegal consonant clusters: acase of perceptual assimilation? Journal ofExperimental Psychology: Human Perception andPerformance 24. 592-608.

Hayes, B. 1995 Metrical Stress Theory. Chicago: TheUniversity of Chicago Press.

Jusczyk, P. 1999 Dividing and conquering linguisticinput. In M. Gruber, K. Olson and T. Wysocki,eds. Proceedings of the Chicago Linguistic Society34. Volume II: The Panels. Chicago: ChicagoLinguistic Society.

Jusczyk, P. and R. Aslin 1995 Infants detection ofsound patterns of words in fluent speech.Cognitive Psychology 29. 1-23.

Jusczyk, P., A. Cutler, and N. Redantz 1993 Preferencefor the predominant stress pattern of Englishwords. Child Development 64. 675-687.

Jusczyk, P., A. Friederici, J. Wessels, V, Svenkerud, andA. Jusczyk 1993 Infants sensitivity to the soundpattern of native language words. Journal ofMemory and Language 32. 402-420.

Jusczyk, P., Luce, P., and Charles-Luce, J. 1994 Infants'sensitivity to phonotactic patterns in the nativelanguage. Journal of Memory and Language 33,630-645.

Kuhl, P., J. Andruski, I. Chistovich, L. Chistovich, E.Kozhevnikova, V. Ryskina, E. Stolyarova, U.Sundberg, and F. Lacerda 1997 Cross-languageanalysis of phonetic units in language addressed toinfants. Science 277(5326). 684-6.

Kuhl, P., K. Williams, F. Lacerda, K. Stevens, and B.Lindblom 1992 Linguistic experience altersphonetic perception in infants by six months ofage. Science 255. 606-608.

Lehiste, I. 1970 Suprasegmentals. Cambridge, MA:MIT Press.

Mandel, D., P. Jusczyk, and D. Pisoni 1995 Infantsrecognition of the sound pattern of their ownnames. Psychological Science 6. 314-317.

Massaro, D. and M. Cohen 1983 Phonological contextin speech perception. Perception andPsychophysics 34. 338-348.

Matteson, E. 1965 The Piro Arawakan Language.Berkeley and Los Angeles: University ofCalifornia Press.

Mehler, J., J. Bertoncini, E. Dupoux, and C. Pallier1996 The role of suprasegmentals in speechperception and acquisition. In T. Otake and A.Cutler, eds. Phonological structure and languageprocessing. Cross-linguistic studies. 145-169.Berlin: Mouton de Gruyter.

Mehler, J., E. Dupoux, and J. Segui 1990 Constrainingmodels of lexical access: The onset of wordrecognition. In G. Altman, ed. Cognitive models ofspeech processing: Psycholinguistic andcomputational perspectives. 236-262. Cambridge,MA: MIT Press.

Mehler, J., P. Jusczyk, G. Lambertz, N. Halsted, J.Bertoncini, and C. Amiel-Tison 1988 A precursorof language acquisition in young infants.Cognition 29. 144-178.

Moon, C., R. Cooper, and W. Fifer 1993 Two-day-oldsprefer their native language. Infant behavior anddevelopment 16. 495-500.

Morgan, J., R. Shi, and P. Allopenna 1996 Perceptualbases of rudimentary grammatical categories:Toward a broader conceptualization ofbootstrapping. In: J. Morgan and K. Demuth eds.Signal to Syntax. Bootstrapping from Speech toGrammar in Early Acquisition. Mahwah, N.J.:LEA. 263-283.

Ntnen, R., A. Lehtokovski, M. Lennes, M. Cheour,M. Huotilainen, A. Iivonen, M. Vainio, P. Alku, R.Ilmoniemi, A. Luuk, J. Allik, J. Sinkkonen, and K.Alho 1997 Language-specific phonemerepresentations revealed by electric and magneticbrain responses. Nature, 385. 432-434.

Navarro Tomas, T. 1965 Manual de PronunciacinEspaola. Madrid: Consejo Superior deInvestigaciones Cientficas.

Nazzi, T., J. Bertoncini, and J. Mehler 1998 Languagediscrimination by newborns: toward anunderstanding of the role of rhythm. Journal ofExperimental Psychology: Human Perception andPerformance 24. 756-66.

Nespor, M. and I. Vogel 1986 Prosodic Phonology.Dordrecht: Foris.

Neville, H., D. Mills, and D. Lawson 1992 Fractionatinglanguage: Different neural subsystems withdifferent sensitive periods. Cerebral Cortex 2.244-258.

Pallier, C., L. Bosch, and N. Sebastin-Galls 1997 Alimit on behavioral plasticity in speech perception.Cognition 64. B9-B17.

Peperkamp, S. 1997 Prosodic words. The Hague:Holland Academic Graphics.

Peperkamp, S., and E. Dupoux submitted On infantsacquisition of phonological alternations. Submittedto: I. Lasser ed. Proceedings of GALA. Peter LangVerlag.


Peperkamp, S., E. Dupoux and N. Sebastin-Galls1999 Perception of stress by French, Spanish, andbilingual subjects. Proceedings of EuroSpeech99, Vol. 6. 2683-2686.

Pitt, M. 1998 Phonological processes and the perceptionof phonotactically illegal consonant clusters.Perception and Psychophysics 60. 941-951.

Polivanov, E. 1974 The subjective nature of theperceptions of language sounds. In E. Polivanov.Selected Works. Articles on General Linguistics.Compiled by A. Leontev. 223-237. The Hague:Mouton

Polka, L. 1991. Cross-language speech perception inadults: Phonemic, phonetic and acousticcontributions. Journal of the Acoustical Society ofAmerica 89 (6). 2961-2977.

Polka, L. and J. Werker 1994 Developmental changes inperception of non-native vowel contrasts. Journalof Experimental Psychology: Human Perceptionand Performance 20. 421-435.

Rice, K. 1987 On defining the intonation phrase:evidence from Slave. Phonology Yearbook 4. 37-59.

Rietveld, A. 1980 Word boundaries in the Frenchlanguage. Language and Speech 23. 289-296.

Sapir, E. 1921 Language. New York: Harcourt BraceJovanovich.

Shady, M., P. Jusczyk, and L. Gerken 1998 Infantssensitivity to function morphemes. Paper presentedat the 23td Annual Boston University Conferenceon Language Development, Boston, MA.

Schane, S. 1968 French Phonology and Morphology.Cambridge, MA: MIT Press.

Shi, R. 1995 Perceptual Correlates of Content Wordsand Function Words in Early Language Input,unpublished Ph.D. diss., Brown University.

Shi, R., J. Morgan, and P. Allopenna 1998 Phonologicaland acoustic bases for earliest grammaticalcategory assignment: a cross-linguisticperspective. Journal of Child Language 25. 169-201.

Schtz, A. 1985 The Fijian Language. Honolulu:University of Hawaii Press.

Trubetzkoy, N. 1939 Grundzge der Phonologie.Travaux du Cercle Linguistique de Prague 7.

Vago, R. 1980 The Sound Pattern of Hungarian.Washington, D.C.: Georgetown University Press.

Vitevitch, M., P. Luce, J. Charles-Luce, and D.Kemmerer 1997. Phonotactics and syllable stress:implications for the processing of spoken nonsensewords. Language and Speech 40. 47-62.

Vogel, I. 1982 La sillaba come unit fonologica.Bologna: Zanichelli.

Vogel, I. 1988 Prosodic constituents in Hungarian. InP.M. Bertinetto and M. Loporcaro, eds. CertamenPhonologicum. Papers from the 1987 CortonaPhonology Meeting. 231-250. Tornino: Rosenberg& Sellier.

Werker, J. and R. Tees 1984a Cross language speechperception: Evidence for perceptual reorganizationduring the first year of life. Infant Behavior andDevelopment 7. 49-63.

Werker, J. and R. Tees 1984b Phonemic and phoneticfactors in adult cross-language speech perception.Journal of the Acoustical Society of America 75.1866-1878.

Wiese, R. 1996 The phonology of German. Oxford:Oxford University Press.

Documents

Fossil Markers of Language Development Phonological_deafness.in_royaumont_OUP