JACQUES MEHLER, JEAN YVES DOMMERGUES, AND Uu FRAUENFELDER

CNRS

In this study a monitoring technique was employed to examine the role of the syllable inthe perceptual segmentation of words. Pairs of words sharing the first three phonemes buthaving different syllabic structure (for instance, pa-/ace and pa/-mier) were used. The targetswere the sequences composed of either the first two or three phonemes of the word (forinstance, pa and pal). The results showed that reaction times to targets which correspond tothe first syllable of the word were faster than those that did not, independently of the targetsize. In a second experiment, two target types, V and VC (for instance, a and a/in the twotarget words above) were used with the same experimental list as in experiment one. Sub-jects detected the VC target type faster when it belonged to the first syllable than when itbelonged to the first two syllables. No differences were observed for the V target type whichwas in the first syllable in both cases. On the basis of the reported results an interpretation inwhich the syllable is considered a processing unit in speech perception is advanced.

The perceptual units into which the con-tinuous speech signal is segmented con-stitute one of the central issues in psycho-linguistics. Investigators have argued for theperceptual primacy of a wide variety ofcandidates such as the phoneme, syllable,word, and sentoid. Traditionally, thephoneme has been taken as the basic per-ceptual unit. However, the lack of invari-ance between the acoustic signal and thephoneme has weakened its candidacy. In-deed, the acoustic information correspond-ing to a phoneme is often distributedthroughout several segments or inversely,the information about several phonemes isconcentrated into one short stretch ofacoustic signal (Liberman, 1970).

Such facts as these have led some to

The authors wish to express their thanks to YvesBarbin for his invaluable technical assistance. Thiswork was supported by grants from Harry FrankGuggenheim Foundation 1977/1978 and D.G.R.S.T.1980 to J. Mehler. Requests for reprints should besent to Dr. J. Mehler, Laboratoire de Psychologie,CNRS, 54, Boulevard Raspail, 75006 Paris.

0022- 53711811030298-08$02.00/0Copyright © 1981 by Academic Press, Inc.All rights of reproduction in any form reserved.

abandon the phoneme and endorse the syl-lable as the basic segmentation unit. Ac-cordingly, some device might operate au-tomatically on the speech wave segmentingit into syllabic units. The feasibility of sucha view is corroborated by the frequent in-corporation of the syllable into speech rec-ognition systems (Mermelstein, 1975; Vais-siere, 1980). Further evidence in support ofthe syllable has also accumulated in otherdomains. By means of a recognition mask-ing technique, Massaro (1974) has shownthe existence of a 250-millisecond process-ing and storage period which defines a unitcorresponding to the syllable.

There are also some data bearing on theclassificatory status of the syllable in youngchildren and adults. Liberman, Shank-weiler, Fischer, and Carter (1974) havetested young children on a tapping game.They showed that children (4- 5 yearsold) could identify by the number of tapssyllabic, but not phonetic, segments; onlyolder children (6-7 years old) were ableto perform on both tasks. Inspired by

these results, Morais, Cary, Alegria, andBertelson (1979) explored whether thisperformance difference is due to cognitivegrowth or to some explicit training proce-dure (learning to read and write usually oc-curs between ages 4 and 7). They haveshown that illiterate adults perform more orless like the younger children. Neither canadd phonemes to or delete them from wordsor nonwords. However, adults who hadonly recently learned the phoneme-graph-eme correspondence system were ableto perform on both phoneme and sylla-ble tasks. These results suggest that thesyllable is a unit available for classificatorypurposes early, whereas the phoneme isonly available to those having mastered thephoneme-grapheme correspondence.

Some more direct evidence concerningthe role of the syllable in speech recognitioncomes from monitoring studies. In a studyconducted to test the perceptual reality ofphonemes and syllables (Savin & Bever,1970), subjects were asked to detect asquickly as possible either the first phonemeof a syllable or the syllable itself presentedin a list of nonsense syllables. The system-atically shorter RTs obtained for syllablesled Savin and Bever to conclude thatphonemes are not perceived directly but arederived from an analysis of the syllabic per-ceptual unit.

Both the methodology and the interpre-tation in this experiment have come under acertain amount of attack. Foss and Swinney(1973) propose an alternative explanation inwhich an important distinction is madebetween perception and identification.They maintain that, generally, "smallerunits are identified by fractionating largerones" (p. 254). This hypothesis was ex-tended to predict that the detection ofphonemes and syllables implies the identifi-cation of the word in which they are found.

Initial empirical support for this predic-tion was provided in a study by Rubin, Tur-vey, and Van Gelder (1976) who found thatinitial stop consonants were detected fasterin words than in nonwords. However, these

results were later attributed to the use of amodified monitoring procedure in whichsubjects had to monitor two targets simul-taneously. The abnormally long RTs ob-tained with this procedure suggest that itinduced a lexically based response. Morerecent research has shown that the detec-tion of item initial phonemes and syllablescan be based on a prelexical "phonetic"code (Foss & Blank, 1980; Segui, Frauen-felder, & Mehler, note 1) suggesting thatthe perception or identification of a linguis-tic unit at level n is not necessarily derivedfrom a higher level n + 1.

Many of the other studies, critical ofSavin and Bever (McNeill & Lindig, 1973;Healy & Cutting, 1976; Swinney & Prather,1980; Mills, 1980) suffer from some seriousmethodological weaknesses (see Mehler,Segui, & Frauenfelder, 1980 for more de-tails). Thus, for example, the phonemeconditions used in an alleged comparisonbetween phoneme and syllable detectiontimes (Swinney & Prather: one vowel con-dition; Mills: a match condition) are mostlikely syllable conditions. Finally, none ofthese alternative explanations can explainthe strong correlation between RTs tophonemes and to the syllable in which theyare found (Segui et al., note 1); a result con-sistent with the original hypothesis pro-posed by Savin and Bever.

To examine the role of the syllable inspeech segmentation further, RTs to se-quences of phonemes which did or did notcorrespond to the syllabic structure of thestimulus words were compared. Thus, forexample, although the words palace andpalmier share the first phonemes /p/ /a! IV,they have different syllabic structures: pa... and pal ... , respectively. Subjectswere given target phoneme sequences (paand pal) to detect in stimulus words of bothsyllabic structures. If the stimulus wordsare segmented according to their syllabicstructure, RTs should be faster when thetarget phoneme sequence matches the firstsegmented syllable of the stimulus word.On the other hand, two alternative pho-

neme based predictions can be made byhypothesizing a mechanism not based onsyllabic structure but on the size of thetarget. First, if the subject conducts aphoneme-by-phoneme analysis of thestimulus, shorter RTs could be expected forthe shorter target sequences (RT: pa <pal)since less of the stimulus word (one fewerphoneme) must be analyzed to initiate a re-sponse. Finally the inverse results (RT: pa> pal) is predicted by the uncertainty hy-pothesis (Foss & Swinney, 1973; Swinney& Prather, 1980) according to which RTsdepend on the uncertainty of the subjectsconcerning the nature of the target/stimulus. The smaller the target (in numberof phonemes?), the fewer cues there areavailable for identifying the stimulus andhence the longer the RTs.

These three hypotheses will be put to testin the following experiment.

MethodSubjects. Forty-two subjects (divided

into two groups of 21 each), all nativeFrench speakers from the Parisian uni-versity community, participated in the ex-periment which lasted about one half hour.

Materials and design. Five pairs ofmono morphemic bisyllabic French nounsof similar frequency sharing the same initialthree phonemes (CYe) were selected suchthat these phonemes made up the first syl-lable for one member of the pair and thefirst two (CY) phonemes formed the firstsyllable of the second member of the pair.For instance, in the pair: palace/palmierthe first three phonemes (lp//a1/II) are iden-tical. Yet, this eye sequence correspondsto the first syllable only the word palmier,ey being the first syllable of the wordpalace. For each of the five pairs the initialconsonant was either a voiced or a voice-less stop and the second was a liquid (eitherII/ or Ir/). The five pairs were: palace-palmier, carotte -carton, tarif-tartine,garage -gardien, balance -balcon.

These 10 stimulus words containing thetarget were each placed as the last item inexperimental sequences. Each experimen-tal sequence was composed of the targetitem and 1 to 4 bisyllabic filler words. Bothmembers of a pair appeared in the same po-sition (from the second to the fifth position)in their respective sequences. In addition tothis first set of 10 target sequences, a sec-ond set was used with different fillers butwith the same target words in the same po-sition as in the first set. Each set of targetsequences was mixed with 10 different dis-tractor sequences forming two blocks (A &B), each containing 20 sequences. Distrac-tor sequences either had target words as thelast item (any position from the first to thesixth) or no target to prevent the subjectsfrom anticipating last items in long se-quences. These 40 sequences along with 5warm-ups were recorded by a French na-tive speaker at a normal rate with a two-track Ampex AG440B. The words in eachsequence were separated by 2-secondintervals; sequences were separated by 10seconds.

Table 1 provides a description of the ex-perimental sequences and of the targetsgiven to the two groups of subjects. Twogroups were used to counterbalance thepresentation order of a given stimulus wordand its two corresponding targets.

As can be seen in Table 1, the stimulusword balance in position 2 is associatedwith the target /ball for group 1 and with/bal for group 2 in sequence number 6 ofBlock A. The inverse matching is found insequence number 19 of Block B. Thetargets to be detected were displayed visu-ally on small 3 x 5 inch cards numberedfrom 1to 45. Before hearing each sequence,subjects heard the instruction "next card"and had 10 seconds to turn and read thenext card. Subjects in groups 1 and 2 re-ceived different decks of cards.

The list was presented binaurally andsubjects responded by pressing a responsebutton that stopped an electronic clock in aPDP-12 computer. A click, aligned manu-

TABLE 1DESCRIPTION OF EXPERIMENTAL SEQUENCES AND TARGETS

The position of the stimulus words inTarget for Target forNumber of the experimental sequences

experimental subjects in subjects insequence 2 3 4 Group 1 Group 2

Block A6 MORCEAU BALANCE BAL BA8 JEUDI MUSEE CAR 0 TTE CA CAR17 FICHIER BALLON BA BAL25 SERVICE CHEVEUX CARTON CAR CA

Block B19 MONSIEUR BALANCE BA BAL33 EPOQUE MAISON CAROTTE CAR CA40 PROBLEME CHEVAL CARTON CA CAR42 JOURNEE BALCON BAL BA

ally with the beginning of the target word,triggered the clock. A correction for eachclick was obtained by means of a two chan-nel oscilloscope and was edited into thedata collection program.

the calculation of the means. The omitteddata made up less than 3% of the subjects'responses.

The results obtained in this experimentare displayed in Figure 1. These resultsshow that when the subjects have to re-spond to a ev target in a word whose sylla-ble structure is ev/ ... (hereafter referredto as ev words) their RTs are faster thanwhen they respond to the same word with aeve target (352 msec vs 371 msec, respec-

RESULTS

The mean reaction times for each subjectand for each item were computed. RTslonger than 1000 milliseconds and shorterthan 100 milliseconds were omitted from

380

375

--;;; 370"0

"0u 365OJ

'"....-< 360.-<...-.5

OJ 3558...'-'

" 3500...'-'u 345'"OJ<>:

tively). When responding to a eve wordwith a eve target, subjects respond fasterthan when responding to the same wordwith a ev target (356 msec vs 378 msec,respectively). Since the results for Group 1and Group 2 are analogous, a two-waywithin-subject analysis of variance wasconducted using the combined data. Thetwo main factors were target type (eV vseVe) and word type (eV words vs evewords). The interaction between thesefactors is highly significant, F(l,42) =

11.03, p < .005, but individually they failedto yield any significant difference, F <1 in both cases. Likewise, a two-way anal-ysis of variance was carried out on themean reaction times corresponding to thefive pairs of words giving F(l,4) = 6.75, p <.10 for the interaction between the twomain factors. As before, individually thesefactors were not significant.

A t test was used to compare the meanRTs for both word types as a function of thetarget type. Using the subjects as a randomvariable, a significant difference was ob-tained for both word types; for ev words,t(41) = 2.02, p < .02 (one tailed), for evewords, t(41) = 2.59, p < .01. A t test usingitems as the random variable gave analo-gous differences for ev words, t(4) = 2.67,p< .05, for eve words, t(4) = 2.36,p< .05.

DISCUSSION

These results show that subjects detect atarget phoneme sequence faster when itcorresponds to the first syllable of thestimulus word than when it does not. Thus,the RTs do not depend on target size aspredicted by the two alternative hypothesesbut rather on the syllabic relationship be-tween the target and the stimulus word. Thefollowing (albeit simplistic and schematic)description of what subjects are likely to doin this task provides insight into these re-sults. First, upon seeing the target, subjectsmust form and store some representation ofit. On subsequently hearing the acousticstimulus, they must segment and analyze itto produce a stimulus representation. The

detection process can thus be seen as thematching of stimulus and target repre-sentations.

The analysis of the acoustic signal corre-sponding to a given target word (palmier) isunlikely to vary as a function of the target(/pa/, /pall) given to the subject. The ob-served RT differences can then be attrib-uted to the process of matching between thestimulus and target representations and notto the computation of the former. When thetarget sequence corresponds exactly to thefirst syllable ofthe stimulus word, there is abetter match between the two repre-sentations than when it does not. The vary-ing degree of compatibility between thetarget and the stimulus representation couldthus explain the direction of the RT differ-ences. Since the representation of a giventarget (pa) presumably does not vary, theprocessing of the two words in a pair (pal-mier, palace) must differ. Thus there mustbe acoustic cues in the signal that the sub-jects exploit to derive different stimulusrepresentations for the two words. This in-terpretation can explain the highly signifi-cant interaction found between word andtarget types.

This account is compatible with thetarget - stimulus mismatch hypothesis pro-posed by Mills (l980a; 1980b). Mills arguesthat the closer the match is between thesubjects' expectancy about the stimulusand the actual acoustic stimulus, the fasterthe subjects respond to the stimulus. Itshould be noted, however, that this expla-nation can not account for the RT differ-ences found here without appealing to asyllabic segmentation unit. Indeed, atarget-stimulus mismatch hypothesis,formulated in phonemic terms would notpredict the observed interaction. If thesubjects' uncertainty concerning the stim-ulus is based only on the number of pho-nemes in the target, then longer targetsshould have led to shorter RTs. Thus, asyllable-based hypothesis is necessary.Unfortunately, however, our experimentdoes not allow us to make precise claims

concerning the target representation andthe more-or-Iess analyzed character of thesyllable.

In the hope of better understanding therole of the syllable and its potential con-stituents in word processing, a second ex-periment was carried out. This experimentwas based on the assumption that if sub-jects segment the signal syllabically, thenthey would respond faster to a sequence asfall when it is contained in the same syllablepaLmier than when it is found in two differ-ent syllables paLace. This prediction is de-rived from the general hypothesis accordingto which subjects respond faster to stimulusitems when these belong to the same ratherthan to different constituents at any lin-guistic level (Fodor, Bever, & Garrett,1974).

Method

Subjects. Twenty subjects from the Pari-sian university community participated inthis experiment.

MateriaLs and design. The experimentaldesign and the linguistic materials were thesame as those used in the first experiment.Two groups of 10 subjects received thesame instructions with the exception thatthey were told that the visually specifiedtargets V and ve could appear anywhere inthe word and not just in the initial positionas in the first experiment. For the experi-mental words, the V target corresponded tothe vowel in the first syllable and the vetarget to the first ve sequence in the word.

RESULTS

The mean reaction times for each subjectand for each item were computed. Reactiontimes longer than 1500 milliseconds andshorter than 100 milliseconds were omittedfrom these calculations. The excluded datamade up 6.5% of the subjects' responses.Table 2 shows the overall reaction times.

Table 2 shows that subjects' RTs to vetargets in ev words are significantly slowerthan those to the same target in eve

TABLE 2MEAN REACTION TIMES (msec) FOR ev AND

eve WORDS AS A FUNCTION OF TARGET TYPE

evword

eveword

words, t(9) = I.76, p < .05 (one tailed). Onthe other hand, no differences were foundbetween RTs to V targets in ev and evewords.

DISCUSSION

The results observed for VC targets sup-port our hypothesis, according to whichfaster RTs are expected when the target se-quence belongs to one syllable rather thanto two different syllables. Furthermore, nodifferences were obtained for the V targetin a ev or a eve syllable. These two re-sults taken together are compatible with thesyllabic segmentation hypothesis. How-ever, the longer overall RTs of the order of650 milliseconds do not eliminate the possi-bility of a postlexical access response. Anexperiment by Marslen- Wilson (note 2) andsome preliminary results obtained in ourlaboratory suggest that the task of mon-itoring targets anywhere in a word maylead the subject to rely more heavily on thepostaccess code. If this is the case, theseresults may be taken as an indication thatthe post lexical code (phonological code) isalso syllabic in structure.

GENERAL DISCUSSION

Results of experiments I and II are com-patible with the general hypothesis that thesyllable constitutes a unit of speech process-ing. Indeed, through a process of segmen-tation, subjects seem able to attainsyllable-like constituents for the processingof words and sentences. However, one ofthe major problems with a syllabic parserwas raised by Liberman and Studdert-Kennedy (1978). These authors claim that"syllable boundaries in fluent speech are

frequently random with respect to words ormorphemes" (p. 173). They challengedproponents of syllabic segmentation to ac-count for the way in which the sentence"He's a repeated offender" is parsed syl-labically in a way compatible with its mor-phological structure. A partial solution tothis challenge may come from the signal it-self and/or from top-down constraints. Re-cent results reported by Mills (1980) suggestthat word boundary information is coded inthe syllable. Mills showed that a evetarget sequence which corresponds to aword (can) was detected faster in the samemonosyllabic word than either in this wordspliced out of bisyllabic and trisyllabicwords (candle, candlelight) or in thesewords themselves. Thus the word boundarycues in syllables would improve the map-ping of syllable on the morphemes andwords. Furthermore, top-down constraintsmay also help with this mapping process.

The evidence for syllabic segmentationpresented here clearly has its implicationsfor lexical access. As was pointed out ear-lier, the subjects' detection response prob-ably precedes lexical access and thus isbased on the prelexical code (Foss &Blank, 1980). Accordingly, the postulatedsyllabic segments could well serve as ac-cessing units. This claim is at odds withlexical access models based on phoneticunits. For these models, the words palmierand palace are still potential word candi-dates (in the same cohort) for the subjecthaving heard the sequence /pall. In the syl-labic hypothesis, these two words could bedistinguished earlier because their syllabicstructure furnishes more information thanwas usually assumed.

The results reported here provide evi-dence for syllabic segmentation of speech.However, more research is necessary todetermine which acoustic cues are actuallyused in syllabic segmentation and also inlexical access.

REFERENCES

FODOR, J. A., BEVER, T. G., & GARRETT, M. F. The

psychology of language. New York: McGraw-HiIl,1974.

Foss, D. S., & BLANK, M. A. Identifying the speechcodes. Cognitive Psychology, 1980, 12, 1-31.

Foss, D. S., & SWINNEY, D. A. On the psychologicalreality of the phoneme: Perception identificationand consciousness. Journal of Verbal Learningand Verbal Behavior, 1973, 12,246-257.

HEALY, A. F., & CUTTING, J. E. Units of speech per-ception: Phoneme and syllable. Journal of VerbalLearning and Verbal Behavior, 1976, 15,73-83.

LIBERMAN, A. M. The grammars of speech and lan-guage. Cognitive Psychology, 1970, 1,301 -323.

LIBERMAN, A. M., & STUDDERT-KENNEDY, M.Phonetic perception. In R. Held, H. W.Leibowicz, & H. L. Teuber (Eds.), Handbook ofsensory physiology. VII. Perception. Berlin:Springer- Verlag, 1978.

LIBERMAN, I. Y., SHANKWEILER, D., FISCHER,F. W., & CARTER, B. Reading and the awarenessof linguistic segments. Journal of ExperimentalChild Psychology, 1974, 18,201 -212.

MASSARO, D. Perceptual units in speech recognition.Journal of Experimental Psychology, 1974,102(2), 199-208.

McNEILL, D., & LINDIG, K. The perceptual reality ofphonemes, syllables, words and sentences. Jour-nal of Verbal Learning and Verbal Behavior,1973, 12, 419-430.

MEHLER, J., SEGUI, J., & FRAUENFELDER, U. Therole of the syllable in language acquisition andperception. In T. F. Myers, J. Laver, & J. Ander-son (Eds.), The cognitive representations ofspeech. Advances in Psychology series.Amsterdam/New York: North-Holland, 1980.

MERMELSTEIN, P. Automatic segmentation of speechinto syllabic units. Journal of Acoustical Societyof America, 1975, 58(4), 880-883.

MILLS, C. B. Effects of context on reaction time tophonemes. Journal of Verbal Learning and Ver-bal Behavior, 1980a, 19,75-83.

MILLS, C. B. Effects of the match between listenerexpectancies and coarticulatory cues on the per-ception of speech. Journal of Experimental Psy-chology: Human Perception and Performance,1980, 6(3), 528- 535.

MORAIS, J., CARY, L., ALEGRIA, J., & BERTELSON, P.Does awareness of speech as a sequence ofphones arise spontaneously? Cognition, 1979, 7,323-331.

RUBIN, P., TURVEY, M. T., & VAN GELDER, P. Initialphonemes are detected faster in spoken wordsthan in spoken nonwords. Perception andPsychophysics, 1976, 19(5),394-398.

SAVIN, H. B., & BEVER, T. G. The nonperceptual re-ality of the phoneme. Journal of Verbal Learningand Verbal Behavior, 1970,9,295-302.

SWINNEY, D. A., & PRATHER, P. Phonemic identifica-tion in a phoneme monitoring experiment: The

variable role of uncertainty about vowel contexts.Perception and Psychophysics, 1980, 27(2),104-110.

VAISSIERE, J. Speech recognition as models of speechperception. In T. Myers, J. Laver, & J. Anderson(Eds.), The cognitive representation oj speech.Advances in Psychology series. Amsterdam/NewYork: North-Holland, 1980.

REFERENCE NOTES

I. SEOUl, J., FRAUEN FELDER, U., & MEHLER, J.Phoneme monitoring, syllable monitoring andlexical access, in preparation.

2. MARSLEN-WILSON, W. D. Sequential decisionprocess during lVord recognition. Paper presentedat the Psychonomic Society meetings in San An-tonio, Texas, November 1978.

(Received August 27, 1980)