Full text of DOI:10.3758/BF03196435

Memory & Cognition
2002, 30 (5), 798-810
Perception of resyllabification in French
M. GARETH GASKELL
University of York, York, England
ELSA SPINELLI
Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands
and
FANNY MEUNIER
MRC Cognition and Brain Sciences Unit, Cambridge, England
In three experiments, we examined the effectsof phonological resyllabificationprocesses on the per-
ception of French speech. Enchainment involvesthe resyllabificationof a word-final consonant across
a syllable boundary (e.g., in
, the / / crosses the syllable boundary to become syllable-
chaque avion
k
initial). Liaison involves a further process of realizationof a
consonant, alongside resyllabifica-
latent
). If the syllable is a dominant unit of perception in French (Mehler, Dom-
petit avion
tion (e.g.,the / / in
t
mergues, Frauenfelder, & Segui, 1981), these processes should cause problems for recognition of the
following word. A cross-modal priming experiment showed no cost attached to either type of resyl-
labification in terms of reduced activation of the following word. Furthermore, word- and sequence-
monitoring experiments again showed no cost and suggested that the recognition of vowel-initial words
may be facilitatedwhen they are preceded by a word that had undergone resyllabificationthrough en-
chainment or liaison. We examine the sources of information that could underpin facilitationand pro-
pose a refinement of the syllable’s role in the perception of French speech.
Languagessuchas French, Spanish, and Portuguesehave the target matched the first syllable of the stimulus word
balance
a clear, simple syllabicstructure. Consequently, for the lan- entirely. That is, responses to the /ba/–
and
pairings were relatively fast, as compared
balance balcon
balcon
guage learner and the adult native speaker of such a lan- /bal/–
guage,thesyllableis an importantperceptualunit.In a clas- with the /bal/–
sic study, Mehler, Dommergues, Frauenfelder, and Segui In subsequent research within and between a variety of
and /ba/–
pairings.
(1981) demonstrated the importance of syllabic informa- languages, the generalityof this effect has been examined.
tion,usinga syllable-monitoringtask. NativeFrench speak- StudiesofSpanishandPortugueserevealedsimilarstimulus–
balcon
balance
.
ers listened to pairs of words, such as
and
target interactions,whereas English, which has a less reg-
These bisyllabicitems differ in terms of the positionof the ular syllabic structure, failed to show the critical interac-
syllable boundary with respect to the first three segments. tion (Bradley, Sánchez-Casas, & García-Albea, 1993; Cut-
bal con
(syllable boundaries are marked with a full ler, Mehler, Norris, & Segui, 1986;Morais, Content, Cary,
In
.
stop), the boundary occurs—as it would in an equivalent Mehler, & Segui, 1989). However, the syllable effect in
English word—between the /l/ and the /c/. However, in Spanish appears to rely on relatively slow responses
ba lance
.
the boundary occurs between the /a/ and the /l/. (Sebastian-Gallès, Dupoux, Segui, & Mehler, 1992), and
This is different from the equivalent case in English, in syllable effects in Catalan occur only in certain circum-
which the /l/ is ambisyllabic. Mehler et al. showed that stances (Sebastian-Gallès et al., 1992). The conclusion
when participants were asked to monitor for syllable se- drawn from this body of research is that syllable regular-
quences such as /ba/ or /bal/, responses were faster when ity and transparency are both importantdeterminers of the
reliance of the perceptualsystem on the syllable.In stress-
timed languages like English, the syllable in itself is not
the most important unit. Instead, a strong syllable (a sylla-
ble containinga full vowel) is a more reliablecue to a word
onset(Cutler& Norris, 1988).For syllable-timedlanguages,
the ease with which subsyllabic units can be identified is
important,with a greater reliance on syllableunits in a lan-
guagewith many vowels (e.g., French) and less reliance on
syllables in a language with a smaller vowel inventory
(e.g., Spanish).1
Part of this research was carried out while the first two authors were
at the MRC Cognition and Brain Sciences Unit, Cambridge. This arti-
cle is based in part on a talk given at the XXVII International Congress
of Psychology, Stockholm, Sweden, July 2000. We thank Juan Segui for
advice and the use of his facilities in testing participants and members
of the MRC Cognition and Brain Sciences Unit in Cambridge for valu-
able discussions on this research. Thanks are also due Sami Boudelaa,
Nicolas Dumay, Uli Frauenfelder, David Gow, and an anonymous re-
viewer, who commented on an earlier version of this manuscript. Corre-
spondenceconcerning this article should be addressed to M. G. Gaskell,
Department of Psychology, University of York, York YO10 5DD, Eng-
land (e-mail: g.gaskell@psych.york.ac.uk).
These perceptual procedures are grounded in the statis-
tics of the language. Cutler and Carter (1987) showed that
Copyright 2002 Psychonomic Society, Inc.
798

PERCEPTION OF FRENCH RESYLLABIFICATION
799
more than 90% of English content words begin with a not. The same pattern of results was found for offset-
apple vuffapple
strong syllable, and for the remaining items there is a cost embedded words, such as
embedded in
(impossibleword). Norris et al.
known word boundary
fapple
attached, in that they are more likely to be misperceived (possibleword) and
(Cutler & Butterfield, 1992).2 Likewise, other statistical also argued that a
can be general-
strategies have gained favor as developmental and adult ized to include likely word boundaries, where a likely
segmentation models, based on phonotacticand distribu- boundary is a syllable onset. Thus, any misalignment be-
tional properties of speech (e.g., Brent, 1999; Saffran, tween a word onset and a likely word boundary defined in
Aslin, & Newport, 1996).What thenis thecost ofa syllable- this way should have an associated cost, in terms of de-
based model of the perceptionof French? Just as, for Eng- layed recognitionof the misalignedword. Cutler, Demuth,
lish, it is more difficult to recognize content words with and McQueen (2002) argued, on the basis of studies in a
weak–strong metrical stress patterns (Cutler & Butter- varietyof languages,that the PWC is a language-universal
field, 1992), so in French, it may be difficult to recognize constraint irrespective of the size of legal word units.
words whose onsets do not coincidewith a syllablebound- French is an interesting test case for the PWC, because
ary. This can occur for a number of reasons, particularly single consonantscan, in fact, be possiblewords in French
l’eau
[the water]) and because of the preva-
One reason for misalignmentis that many French func- lence of resyllabification. Norris et al. noted that this
tion words are reduced to single consonants when fol- strategy would cause problems for resyllabification and,
for words beginning with vowels.
(e.g., the /l/ in
l’eau t’appelle n’aime
). The words particularly, French liaison. Their suggested solution was
lowed by a vowel (e.g.,
,
,
following these reduced forms are misaligned with sylla- that the PWC might be switched off when a liaison con-
enchainment
ble onset. Resyllabification,or
, of word-final sonant was recognized. The present research tests whether
consonants creates the same state of affairs. For example, this is the case.
chaque
the final consonant of
would be syllable-final
chaque
Vroomen and de Gelder (1997) found evidence for a
syllable–word boundary misalignmentcost, althoughthey
when followed by a consonant such as /t/ (e.g.,
train
[each train]) but would migrate to the onset of the did not interpret their effect in terms of the PWC. They
chaque
following syllablewhen followed by a vowel (e.g.,
used cross-modal priming to examine the activation of
avion
[each aircraft]). Resyllabification can also occur in various types of embedded words in Dutch. When the em-
liaison
, in which a
grates to syllable-onsetpositionwhen followed by a vowel in
or, less frequently, a glide (e.g., the final /t/ in
latent
boos
combinationwith
consonantmi- bedded word matched a syllable onset (e.g.,
[angry]
[raspberry]), the recognitionof a word related
to the embedded word was facilitated. This effect did not
[the little aircraft]). The consonant is termed latent hold when a word was embedded in a longer syllable(e.g.,
framboos
le petit
avion
wijn
zwijn
[swine]). McQueen (1998) reported
because when the word is produced in isolation or is fol-
[wine] in
lowed by a stop consonant or a fricative, the consonant is a comparableeffect—also in Dutch—using word spotting.
le petit chien
rok
not realized (e.g.,
[the little dog]).
He found that a word such as
fiem.rok
(skirt) was easier to de-
(target aligned with syl-
These phenomenaare complex,and theirapplicationde- tect when embedded in
fie.drok
pends on phonetic, syntactic, morphological,and stylistic lable boundary) than when embedded in
(target
factors. Nonetheless,the simple fact remains that there are misalignedwith syllable boundary). A similar but weaker
several related processes that cause glide-initial and, par- effect was found when the word offset was misaligned
ticularly, vowel-initial words to be nonaligned with sylla- with the syllable boundary. In McQueen, these syllable
ble onsets in connectedspeech. Our goal in this study was boundaries were enforced by the phonotactic rules of the
to examine the consequences of this misalignment on the language, suggesting that syllabification effects may
recognition of such words.
partly reflect the use of phonotactic knowledge in speech
Although there are differences between languages in segmentation.For French, Dumay, Frauenfelder, and Con-
terms of the importance of the syllable, there are also tent (2002) studied syllable–word alignment effects at
groundsfor supposingthat misalignmentbetween syllable bothonset and offset, again using word spotting.Like Mc-
and word boundarieswill cause problems for word recog- Queen, they found a processing cost involved in syllable
nitioninalllanguages.The possibleword constraint(PWC), misalignment, although in this case, it was significant
developed by Norris, McQueen, Cutler, and Butterfield only when the onset of the word mismatched the syllable
lac
zun.lac
vs.
(1997), states that a potential segmentation of the speech boundary (e.g.,
stream is disfavored if it leaves a consonant or consonant
sequence stranded between a known word boundary and
[lake] embedded in
zu.glac
).
These studies suggest that words that are misaligned
a hypothesized boundary. This constraint was used to ex- with syllable boundaries are relatively difficult to recog-
plain data from word-spotting experiments, in which par- nize, in both syllable-timed and stress-timed languages.
ticipants were asked to press a button as soon as they de- However, none of the studies mentionedso far tested a sit-
tectedany word in a short sequence of speech. Norris et al. uation in which misalignment was embedded in normal
sea seashub
in
argued that it is easier to spot the word
than sentence context, which is relatively common in French
seash
in
because, although both segmentations involving throughresyllabificationandreductionprocesses.Vroomen
sea
seashub
leave a nonword residue, the residue in
is a and de Gelder (1999) addressed the issue of sentence-
seash sh
shub
), whereas the residue in
possible word (
( ) is embedded resyllabification for Dutch by asking partici-

800
GASKELL, SPINELLI, AND MEUNIER
boot
pantstomonitorfor phonemes,suchasthe/t/ in
(boat). /ta/-initial words may take some time to diminish. These
The phonemes either occurred in their normal syllable- ambiguitieswere studiedby Dejeande la Bâtieand Bradley
de boot die gezonkenis
[the boat,which (1995), who asked participants to monitor for a word-
finalposition(e.g.,
,
is sunk]) or were resyllabified owing to the presence of a initial /t/ in phrases like the ones above. A participant per-
de boot is gezonken
excel-
[the boat is forming correctly on this task would respond when
following vowel (e.g.,
lent tableau
excellent acteur
French speakers made very few errors in this task, but
sunk]). Once again, the misalignment of word and sylla-
ble boundaries resulted in slower performance in the re-
syllabification condition.
(potentialliaison)was presented,but not when
(actual liaison) was presented. Native
Therefore, at least in Dutch, there is a resyllabification there was some evidencethat potentialliaisonsdelayed re-
cost in terms of identifyingthe resyllabified consonantit- sponses. Responses were slower to potential liaison se-
self. If the syllable is a more influential perceptual unit in quences than they were to unambiguous sequences (e.g.,
vrai tableau
French than in languageslike English and Dutch (Mehler
[real painting]). This difference was signifi-
et al., 1981), resyllabification costs may be even more cant (64 msec) when the sentence contexts were fairly
drastic in French. However, Vroomen and de Gelder (1999) short and neutral, but it was not significant(24 msec) when
examined the effects of resyllabification on the recogni- the sentence contexts were fuller and more informative.
tion of the resyllabified consonant itself. Although resyl- Therefore, there is an added reason to suspect that liaison
labification may induce uncertainty about the status of sequencesmay affect the recognitionof the followingseg-
that particular consonant, identification of the resyllabi- ment. On top of the fact that liaison causes misalignment
fied consonantmay have predictive value for the recogni- between syllable and word boundaries, it could also in-
tion of the following word. The situationis similarin cases crease the number of lexical hypothesesthat must be con-
of assimilation of place of articulation, in which the as- sidered, which would further delay the activation of the
similated consonantbecomes somewhat ambiguous(e.g., following word. However, it is important to bear in mind
Gaskell & Marslen-Wilson,1998) but identificationof the that the situation we focus on here differs in an important
assimilated consonant narrows down the possibilities for way from the situation examined by Dejean de la Bâtie
the next segment, because that next segment must be one and Bradley. Their experimentprimarily examined the ef-
that will allowthe assimilationprocess to occurin the first fects of potential liaison on the identification of a word-
excellenttableau
), whereas here we
place. Furthermore, there is some evidence that resyllab- initialconsonant(e.g.,
ified consonants differ from normal syllable-initial con- study the effect of actual liaison on the recognition of the
sonants(Dumay, Content, & Frauenfelder, 1999; Spinelli, following vowel-initial word.
McQueen, & Cutler, in press). If this is the case (the evi-
There are, therefore, several theoreticalreasons for pre-
dence currently available for French is by no means com- dicting that liaison and enchainment will affect speech
pelling),it may be better to thinkof resyllabificationowing perception.We reportthreeexperimentsinwhich theeffects
to liaison and enchainment as a partial or graded process. of these processeson the recognitionof vowel-initialwords
A graded resyllabificationprocess would cause nonalign- were examined.In one condition,the critical vowel-initial
ment between word and syllable boundaries but would word began at a syllable boundary. In two other condi-
leave some acousticmarkings that couldbe used to counter tions, the critical word onset was misaligned with a sylla-
the negative effects of word boundary misalignment or, bleboundary, because eitherenchainmentor liaisoncaused
possibly, even facilitateprocessing.Again, there is a strong the final consonantof the preceding word to resyllabify. In
parallel with English place assimilation, which can be Experiment 1, we used these phrases as prime sentences
complete or partial. For cases of partial assimilation, the in a cross-modal priming experimentdesigned to examine
perceptionof the assimilated segment can facilitaterecog- the lexical activation of the vowel-initial word at two dif-
nition of the following context that licenses the change ferent points during the perception of that word (roughly
(Gow, 2001; Quené, van Rossum, & van Wijck, 1998).
halfway through the word and at the offset of the word). In
In the case of liaison, resyllabification applies to a la- Experiment 2, we looked at the ease of identification of
tent consonant,which may further complicate the process the onset of the word, using a word-monitoring task. Fi-
of constructing the appropriate segmentation of the nally, in Experiment 3, we addressed the critical acoustic-
speech. It is safe to assume that words for which the final phonetic information in these sequences. The ease of
consonant is often not pronounced will be recognizable recognition of the critical word-initial segments on the
without that consonant. Therefore, there is the potential basis of acoustic-phonetic information alone (i.e., when
for increased lexical competition when the liaison conso- normal lexical competitionwas unavailable)was assessed
excel-
nant is encountered. For example, in the sequence
lent acteur
by asking participants to monitor for short sequences of
(/Ek.se.l~a.tak.t r/; excellent actor), a mis- segments in longer nonword sequences excised from the
segmentation may be briefly entertained in which the /t/ original phrases.
excellent tableau
is treated as word initial, as in the phrase
(/Ek.se.l~a.ta.blo/; excellentpainting).In phonemicterms,
these sequences are identical until the end of the /a/, and
given that the incorrect segmentation is favored by align-
EXPERIMENT 1
In Experiment1, cross-modalprimingwas used to assess
ment with the syllable onset, the competition effects of the effects of resyllabification processes on the activation

PERCEPTION OF FRENCH RESYLLABIFICATION
801
of the followingword. In Experiment1A, a typicaltrial in- words were generally unpredictable, with only 35 out of 1,920 con-
tinuations (1.8%) matching the target word and no significant dif-
volved the presentation of an auditory prime sentence to
ferences between conditions. Of the matching continuations, 29 were
the participant,with the visual form of the final word dis-
attributable to just three stimuli with high cloze ratings (un galant
played at the offset of the sentence. The speed of response
homme [a gallant man], un honnete homme [an honest man], and un
hereux evenement [a fortunate event]). Thus, almost all the continu-
to the visual target gave an indication of the degree to
which the lexical representation of that word had been ac-
tivated. In Experiment 1B, the procedure was repeated
with the target presented earlier in the sentence—50 msec
before the isolation point of the critical word—in order to
assess activation earlier in the recognition process.
ations in the spoken stimuli were unpredictable. Given that cloze
predictabilities were low (and so, floor effects could possibly be
masking a systematic plausibility bias), a two-alternative preference
test was also used to test for possible biases in the context sentences.
Forty native French speakers rated the relative plausibility of either
the liaison versus the enchainment primes or the liaison versus the
syllable-aligned primes (20 participants in each test). In both cases,
the two versions of the prime phrase were presented to the partici-
Method
Participants. Eighty-four paid native French speakers of the pant in written form with instructions to rate their relative plausibil-
same dialect took part in this experiment (40 in Experiment 1A and ity on a 9-point scale (extremes of the scale represented a preference
44 in Experiment 1B). The participants either were students at the for one or other of each pair, and the middle of the scale represented
Université René Descartes in Paris or were recruited in Cambridge no preference). The advantage of this method was that it allowed the
through the university and language schools.
different contexts of each critical word to be compared directly. The
Materials and Design. Thirty vowel-initial words with mascu- pretest produced a satisfactory spread of scores, but overall there
line gender, selected from the Trésor de la Langue Française (Imbs, was no systematic semantic bias toward any one of the conditions.
1971), served as the experimental targets. These were embedded in The mean preference scores were 4.9 for liaison versus enchainment
three types of phrasal context to form three related prime conditions
and 4.8 for liaison versus syllable aligned. Neither value deviated
(see Table 1 for examples and the Appendix for the full stimulus list). significantly from the no-preference baseline of 5. The results of the
In the liaison condition, the prime phrase consisted of the article un two pretests confirmed that there were no differences between the
followed by a word containing a latent final consonant (e.g., généreux), spoken stimuli in terms of contextual plausibility or predictability of
followed by the target word (e.g., italien). The vowel beginning the the final words. There was, however, one unavoidable difference be-
target word ensured that the latent consonant (e.g., the /z/ in tween the different context types. The liaison condition exclusively
généreux) was both fully realized and resyllabified to become syl- used article–adjective–noun phrases, whereas the enchainment con-
lable initial. In the enchainment condition, resyllabif ication oc- dition mostly used the more common article–noun–adjective struc-
curred but involved a consonant that was fully realized in all phono-
ture (21 out of 30 items). This difference was due to the relative
logical contexts (e.g., the /z/ in virtuose). With one exception, the scarcity of nouns that undergo liaison and meant that some targets
liaison and enchainment conditions were matched, within each item were used as adjectives in one condition and as nouns in another.
set, on the identity of the resyllabified consonant (/t/, /z/ or /r/). In The syllable-aligned condition was matched in phrase structure with
the syllable-aligned condition, the onset of the target word coincided
with a syllable boundary. This was achieved by choosing a preced-
ing word that ended in a vowel (e.g., chapeau). The three related
prime conditions were compared with a control condition in which
the enchainment condition. These syntactic differences did not lead
to any differences in the predictability of the final word, but they
need to be borne in mind when the results are interpreted.
A set of 120 filler items was constructed in a similar manner to
the same visual target word was paired with an unrelated prime the test set. Of the fillers, 45 used visual targets that were words and
phrase in order to provide a baseline for measuring response facili- were unrelated to the prime phrase. Thirty of these contained words
tation (e.g., for the target word italien, the control prime phrase was that would be resyllabified if followed by a vowel (15 liaison words,
un mysterieux organisme). The control prime began with a vowel and 15 enchainment) but were followed by a consonant (e.g., un auda-
was embedded in either a liaison or an enchainment sequence (there cieux montagnard, where the /z/ in audacieux remained unarticu-
were equal numbers of each type of control, and this contrast had no lated because of the following stop consonant). These potential re-
effect on the results). The average frequency of the targets (which syllabification sentences ensured that resyllabificat ion was not
were also the prime words in three of the four conditions) was 100 predictable in the stimulus set as a whole. A further 15 fillers with
occurrences per million (median, 22). The control prime words had word targets, like the syllable-aligned primes, used a penultimate
a mean frequency of 52 (median, 23).
word ending in a vowel but differed in that the final word began with
The preceding contexts of the related prime word (see Table 1 for a consonant. The remaining 75 filler items used nonword visual tar-
frequencies and durations of these words) were designed to make the gets, meaning that 50% of the targets in the experiments were words.
final word equally plausible and equally predictable in all three con- The prime phrases for these targets were similar to the primes for the
ditions. Predictability was measured with a cloze test, in which the word targets: Thirty contained liaisons or potential liaisons, 30 con-
first two words of the stimulus phrases were given in written form tained enchainments or potential enchainments, and 15 contained a
to 16 native French-speaking participants, who were asked to pro- vowel-final word followed by a consonant. To ensure that a form link
vide plausible one-word continuations. The test showed that final between prime and target was not predictive of a word target, 22 of
Table 1
Example Stimuli for Experiment 1
Context Word
Length (msec)
Context Word Frequency
(mean/median)
Condition
Liaison
Enchainment
Syllable aligned
Control
Spoken Prime Phrase
Visual Target
un généreux italien
un virtuose italien
un chapeau italien
italien
italien
italien
italien
380
447
391
411
230/58
184/31
86/56
un mysterieux organisme
91/54

802
GASKELL, SPINELLI, AND MEUNIER
the nonword targets were superficially similar to the final word of
the prime phrase (e.g., un navire immerge–IMMERSU). The percentage
of trials in which there was a link between prime and target was 29%.
The stimulus phrases were recorded onto DAT by a native French
speaker (the second author) and were transferred digitally to computer
disk. In Experiment 1A, the alignment point for the presentation of
the visual target was the offset of the prime word (at the end of the
prime phrase), in order to assess lexical activation near the end of the
recognition process. In Experiment 1B, the target was placed at a point
at which lexical competition was expected to be in progress. This point
was selected on the basis of a separate gating task (Grosjean, 1996) in-
volving new recordings of the related prime words in isolation. These
recordings were presented in gradually increasing fragment duration
to 11 native French speakers who did not take part in the main exper-
iment. The shortest fragment consisted of the first 150 msec of the
word, and the fragments increased in length by 50 msec each time
until the full word was presented. For each fragment, the participant
had to attempt to identify the word and rate the confidence of his or
her judgment on a 9-point scale. For each participant and each word,
the gate position at which the correct answer was first given and after
which the participant did not change his or her mind was identified,
and the isolation point of each word was calculated by taking the av-
erage of these gate positions across participants. The mean duration
of the isolated words was 643 msec, and the mean isolation point was
348 msec. The target position for Experiment 1B was chosen to be
50 msec before the isolation point for the critical word (on average,
298 msec into the word). In order to map these target points derived
from isolated gated words onto the same words in sentence context,
the target point was transformed into a percentage duration of the
word, and these percentages were mapped onto the words in context.
Thus, if the isolation point less 50 msec turned out to be 40% of the
way through a word, the alignment point for target onset in Experi-
ment 1B was placed 40% of the way through that word in each of the
three prime conditions containing that word. This controlled in a sim-
ple way for the small variations in duration between the tokens of the
target word in these conditions (the overall durations of the prime
words in the three conditions were similar to the durations of the iso-
lated words; mean prime word durations were 635, 627, and 643 msec
for the liaison, enchainment, and syllable-aligned conditions). The
target alignment points were mapped onto the final word of the con-
trol prime phrases, using the same percentages.
Table 2
Results of Experiment 1 (Cross-Modal Priming)
Experiment 1A
Experiment 1B
Condition
Liaison
Enchainment
Syllable aligned
Control
RT
Diff
ER
RT
Diff
ER
540
533
526
591
51
59
65
3.0
3.4
1.6
6.7
583
579
582
624
41
45
42
0.7
1.5
2.0
1.9
Note—RT, reaction time in milliseconds; Diff, control–test difference
(facilitation effect); ER, error rate as a percentage.
ticipantgroupingin the allocationof participantsto exper-
imental lists (for the participantsanalyses) or the test item
grouping in the allocation of items to lists (for the items
analyses). Because it was includedsolelyto reduce the es-
timate of random variation, effects involving the dummy
variable are not reported (Pollatsek & Well, 1995). Raaij-
makers, Schrijnemakers, and Gremmen (1999) recom-
mended that for experimental designs like ours involving
matched sets of items, a significant by-participant effect
is sufficient to reject the null hypothesis.We note this rec-
ommendationhere but report bothby-participant( 1) and
by-item ( 2) values for the sake of completeness.
The results of Experiment1 are summarized in Table 2.
For Experiment1A, the ANOVA revealed a significantef-
F
F
F
5
p ,
F
5
fect of prime type [ 1(3,108) 15.9,
.001; 2(3,78)
.001]. Planned comparisons showed that there
were significant facilitative effects for the liaison
p , p ,
p ,
24.5,
F
5
F
5
[
1(1,36) 14.5,
.01; 2(1,26) 30.7,
.001], en-
F
5
p ,
F 5
chainment [ 1(1,36) 27.3,
.001; 2(1,26) 62.0,
p ,
F
5
p ,
.001], and syllable-aligned [ 1(1,36) 40.7,
F
5
p ,
.001; 2(1,26) 120.4, .001] conditions. However,
there were no differences between these conditions in
terms of the amount of facilitation. An error analysis
F
5
showed a similar effect of prime type [ 1(3,108) 4.2,
p , p ,
Four experimental lists were created, each containing one of the
four stimulus phrases for each experimental item. This ensured that
any 1 participant would hear only one version of each test item. Ex-
perimental items were pseudorandomly ordered, and the lists began
with 10 practice trials.
F
5
.01; 2(3,78) 4.1,
.05], with fewer errors in the
primed conditions than in the control condition.
Experiment 1B, in which the target was placed earlier
to assess activation during the competition, showed the
same overall pattern of results. There was a strong effect
Procedure. The participants were tested individually in a quiet
room. Each prime was presented auditorily at a comfortable sound
level through Sony CD550 headphones. At the end of the auditory
prime (Experiment 1A) or during the course of the final word of the
prime (Experiment 1B), the target was visually displayed in lower-
case at the center of a computer screen for 480 msec. The prime
phrase was not cut off in either subexperiment. The participants had
been informed that the visual target could be either a word or a pseudo-
word and had to perform a lexical decision task on the visual target
by pressing one of two response buttons as accurately and as quickly
as possible. They were required to press the yes button with the fore-
finger of their preferred hands. The reaction times were measured
from the presentation of the visual target on the screen. Stimulus pre-
sentation was computer controlled using the DMDX experiment gen-
erator software developed by K. Forster and J. Forster at the University
of Arizona. The duration of the session was approximately 20 min.
F
5
p ,
.001;
of overall prime type [ 1(3,120)
12.4,
.001]. Priming levels were about one
F
5
p ,
2(3,78) 8.4,
third lower than the priming levels in Experiment 1A, but
once again, there were individuallysignificant facilitation
F
5
p ,
18.1, .001;
effects for the liaison [ 1(1,40)
F
5
p ,
F
5
2(1,26) 14.0,
.01], enchainment[ 1(1,40) 30.0,
p ,
F
5 p ,
.001; 2(1,26) 13.4,
p ,
.01], and syllable-aligned
p ,
F
5
F
5
[
1(1,40) 26.1,
.001; 2(1,26) 13.3,
.01]
conditions.These numerically weaker effects suggestthat
the target word had not been fully activatedby the time the
target was presented, but nonetheless, there was no effect
of the type of related prime on the amount of facilitation
1 in each case). There were no significant ef-
fects in the error analysis.
F
1 and
F ,
(
2
Results
Participant and item analyses of variance (ANOVAs) Discussion
were performed on the data, using the variable prime type
In Experiment 1, we examined the effects of two types
(four levels).A dummy variablerepresentedeitherthe par- of resyllabificationon the perception of following vowel-

PERCEPTION OF FRENCH RESYLLABIFICATION
803
initial words in French. We anticipated that resyllabifica- spersed with 90 filler phrases selected from Experiment 1. For 30 of
these, the chosen target word was the third and final word of the
tion might result in delayed or weakened activation of the
phrase, whereas for the remaining 60, the chosen target was not in
following word, because of the resulting misalignment of
the phrase. This meant that 50% of the trials contained the target
word and syllable boundary. However, the cross-modal
word. In all, 50% of the targets used in this experiment were vowel
initial, with equal proportions of vowel-initial targets in target-
priming results in both subexperimentsshowed no hint of
an effect. Certainly, if resyllabification does impede the
recognitionof the following word, the effect is swiftly off-
set by the segmentalinformationin the speech stream. Ex-
periment 1B demonstrated this most clearly; with the tar-
get presented less than 300 msec into the spoken word,
there was no difference in the amount of priming found in
the three related conditions, suggesting that lexical acti-
vations were equivalent. This result does not sit easily
with the proposition that the syllable is the critical unit of
perception in French.
Althoughthisresult is perhaps surprising, giventhe the-
oreticaland experimentalbackgrounddescribed in the in-
troduction, it is not particularly strong scientifically; all
three test conditions showed significant facilitation, but
present trials and target-absent trials. Some word-monitoring exper-
iments have used target-absent trials in which there was a strong
form similarity between the target and one word from the stimulus
phrase. This encouraged listeners to wait until recognition was more
or less complete before they responded. Here, however, we wanted to
investigate processing early on in recognition, and so we specifically
avoided this type of foil trial.
Three experimental lists were created, each containing one of the
three stimulus phrases for each experimental item. This ensured that
any one participant would hear only one version of each test item.
Experimental items were pseudorandomly ordered, and each partic-
ipant had 10 trials for each condition. The lists began with 10 prac-
tice trials.
Procedure. The participants were tested in small groups in a quiet
room. They were asked to detect a word in a phrase, with the target
word varying from trial to trial. The target was visually displayed in
there was no significant difference between the three re- lowercase at the center of a computer screen for 750 msec, and then
the screen went blank. The spoken phrase was presented auditorily at
lated conditions.It is possiblethat even in Experiment1B,
a comfortable sound level through headphones. The participants were
our target position was too late to catch the interference
instructed to press a response button as soon as they heard the target
created by resyllabification. We could test this by shifting
word or to do nothing if the target was not present. Response laten-
cies were recorded from the marker point at the onset of the target
the target alignment point to earlier in the prime sentence
so as to examine activations at an even earlier point, but
this would run the risk of weakening priming in all condi-
tions to a nonsignificant level. Priming dropped from
around 60 to 40 msec between Experiment 1A and Ex-
periment 1B, presumably because in Experiment 1B, lex-
ical competition in response to the prime had not been
fully resolved by the time the target was presented. Mov-
ing the target even earlier would most likely reduce the
baselinepriminglevelsevenfurther,minimizingthechances
of finding a difference in priming between conditions.In-
stead, in Experiment 2, the immediate effects of resyllab-
ification and liaison were addressed with a different
task—word monitoring(Kilborn & Moss, 1996). This task
offered the prospect of fast responses to the listener’s eval-
uation of, primarily, the beginningof a spoken word (par-
ticularly when no catch trials were used), and so if there
was an early or transient effect of resyllabification, there
would be a good chance that word monitoring would pick
it up.
word.
Results
ANOVAs were conductedon the responsetime and error
data(see Table3), withtargetcontext(liaison,enchainment,
and syllable aligned) entered as a variable. There was a
F
5
p ,
highly significant main effect [ 1(2,72) 8.0,
.01;
.05], showing that the context of the
F
5
p ,
2(2,54) 3.4,
target word affected speed of response. This effect was in-
vestigated further with planned comparisons. Responses
in the liaison and enchainment conditions were signifi-
cantly faster than those in the syllable-aligned condition
F
5
p ,
[liaisonvs. syllablealigned, 1(1,36) 13.3,
.01, and
.05; enchainmentvs. syllablealigned,
p , p ,
F
F
5
5
p ,
2(1,27) 6.0,
1(1,36) 8.7,
F
5
.01, and 2(1,27) 4.6,
.05].
The enchainment and liaison conditions did not differ in
terms of response speed ( 1;
F , F ,
1). The error analy-
F , F ,
1
2
sis showed no effects of target context (
1;
1).
Because of the many constraints operat1ing on st2imulus
selection, we were unable to choose items that were per-
fectly matched on the frequency of the word preceding the
target word. The mean frequencies of these words were
230, 184,and 86 for theliaison,enchainment,and syllable-
aligned conditions, respectively. These differences were
due partly to the presence of a few high-frequency words,
and in terms of median frequency, the conditionswere bet-
ter matched (median values: 58, 31, and 56 for the liaison,
EXPERIMENT 2
In this experiment, trials began with the visual presen-
italien
), and the participants
tationof the target word (e.g.,
were then asked to monitor for the target in the spoken
un généreux italien
).
phrase (e.g.,
Method
Participants. Thirty-nine paid native speakers of French took part
in this experiment. The participants were recruited in Cambridge
through the university and language schools, and none had partici-
pated in Experiment 1.
Table 3
Results of Experiment 2 (Word Monitoring)
Measure
Liaison
Enchainment
Syllable Aligned
Stimuli and Design. The test stimuli consisted of the 30 sets of
related prime sentences used in Experiment 1 (i.e., the liaison, en-
chainment, and syllable-aligned conditions). These were inter-
RT (msec)
Error rate (%)
438
0.8
446
1.3
482
1.0

804
GASKELL, SPINELLI, AND MEUNIER
enchainment, and syllable-aligned conditions, respec- liaison condition over the syllable-aligned condition but
tively). In order to check whether the frequencyof the pre- would then leave unexplained the advantage of the en-
ceding word had any effect on response speed (cf. Foss, chainment condition over the syllable-aligned condition
1969), we examined the by-item correlation between pre- and the lack of an advantage of liaison over enchainment.
ceding word frequency and mean response time. This Syntactic differences between the conditions were, there-
showed no significant correlations either within condi- fore, unable to fully determine the pattern of recognition
tions or with all three conditions pooled.
A similar analysis in which duration of the preceding time advantage in the liaison condition alone.
word was used showed a significant negative correlation
A second potential explanation of the facilitatory ef-
with response time when data were pooled for the three fects found in Experiment 2 relies on successful recogni-
r 5 2 p ,
.05), although the effects tion of the word containing the resyllabified consonant
speed found but may have contributed to the response
conditions (
.219,
within each condition did not reach significance. This ef- prior to encountering the resyllabification itself. If this is
fect showed that response times tendedto be shorter when the case, the recognitionsystem may be able to use the lex-
the word preceding the target was long. Given that word ical information about the form of the recognized word to
offsets will be, on the whole, more identifiable for longer identify the final consonant as being resyllabified and
words, this result may suggest that lexical information word final, rather than simply word initial. For example,
about where a word is likely to end can facilitate recogni- considerthe case in which a participanthears a phrase con-
un virtuose italien
). In many
tion of the following word. Nonetheless, the effect of tar- taining enchainment (e.g.,
get word context cannot be explained by durational dif- cases, the word preceding the target will be recognizable
ferences in the preceding word, because the main effect in before the final consonant. If the final consonant is then
F
5
theresponsetimeanalysisremainedsignificant[ 2(2,53)
recognized as syllable initial, the segment following it
.05] even when the duration of the preceding must be one that licenses resyllabification(i.e., a vowel or
p ,
3.3,
word was included as a covariate in the items analysis. one of a limited set of consonants). Therefore, provided
the word preceding the target can be recognized before
offset, the resyllabified consonant can be used as a cue to
Discussion
Experiment 2 was intended to examine the effects of theupcominglicensingcontext,whichmay facilitaterecog-
misalignment between syllable and word onsets early on nition. The perceptual system has shown a similar sensi-
in the processing of vowel-initial words. The speed of re- tivity to phonological context in the case of alternations
sponses in this experiment(mean response times were be- that were due to place assimilation (Gaskell & Marslen-
tween 400 and 500 msec from word onset)suggeststhat this Wilson, 1998; Gow, 2001).
goal was achieved. By comparison, the priming experi-
It is important to note, however, that this mechanism for
mentswith theearlier target onset (Experiment1B) elicited facilitationis more viable for enchainmentsequencesthan
responsetimes of between 850 and950 msec from the word for liaison sequences, because the presence of the final
onset (or between 550 and 650msec from targetonset). The consonantis obligatoryin enchainment,but not in liaison.
absence of catch trials with stimuli closely related but not This means that the occurrence of a liaison consonant
identical to the target word enabled the participants to re- does not necessarily mark the end of a word. Instead, it
spondon the basis of partiallymatchingonset information. could be the beginning of a new word that coincidentally
excellent acteur
ex-
In these circumstances, we found strong effects of the pre- matches the liaison consonant(cf.
vs.
; Dejean de la Bâtie & Bradley, 1995).
cellent tableau
ceding context of vowel-initial words on their recognition.
Contrary to our expectations, targets preceded by liaison Therefore, on the basis of lexically governed phonologi-
or enchainmentconsonantswere recognized more quickly cal constraints, the target should be less predictable in the
than targets alignedwith syllableboundaries.Instead of li- liaison condition than in the enchainment condition.
aison and enchainment impeding the recognition of the
An alternative or supplementary interpretation relies
following word, it seems that these processes actuallyfacil- more on acoustic markings of resyllabification. Although
itated the recognition of vowel-initial words. Clearly, these French is often cited as a languagein which resyllabifica-
results do not conform to a segmentation strategy for tions are clear, there is some evidence that a resyllabified
French thatplacesgreat emphasison straightforward align- consonantin French is not acousticallyidenticalto a word-
ment between word and syllable onsets, since the only initial consonant. Yersin-Besson and Grosjean (1996)
son
onset-alignedstimuliwere the ones responded to the most looked at the perception of word sequences such as
oeuf
slowly. How then can these findings be explained?
(his egg), which, because of the liaison process, could
son
The liaison condition used an article–adjective–noun be mistaken for an alternative word sequence (e.g.,
neuf
syntactic structure, whereas the other two conditions pre-
[his nine]). With a forced-choice recognition test,
dominantly used an article–noun–adjectivestructure. It is they found that listeners did find these phrases lexically
possible that the article–adjective–noun structure is asso- ambiguous, with only a small preference for the “correct”
ciatedwith facilitatedprocessingof the final word in some disambiguation over the nonliaison alternative. Yersin-
way (although our pretests showed that there was no dif- Besson and Grosjean also examined liaison sequences
ference in terms of semantic plausibility or predictability phoneticallyand found that liaisonconsonantswere, on av-
of the final word). This could explain the advantageof the erage, 10% shorter in duration than nonliaisonconsonants

PERCEPTION OF FRENCH RESYLLABIFICATION
805
(although this difference was based on only eight item
pairs and was not tested for significance).
To summarize, there are four potential sources of infor-
mation that, in some combination,may provide the means
More recent evidencesuggeststhat the completenessof for the facilitation effects shown in Experiment 2. Syntac-
resyllabification involved in enchainment may vary de- ticdifferencesbetween the auditorystimulimay contribute
pending on the consonants involved. Dumay et al. (1999) to the advantage for the liaison condition. Lexical knowl-
showed that there were significantacousticdifferences be- edgeaboutthe offset of the precedingword, combinedwith
tween obstruent–liquid clusters (e.g., /tr/) involved in re- sensitivity to phonological constraints, may facilitate pro-
tante roublarde
same consonants word initially (e.g.,
syllabification (e.g.,
[wily aunt]) and the cessing,particularlyin theenchainmentcondition.Acoustic
troublant
[perturb- markings of resyllabification may facilitate processing in
ing]). This was not the case for resyllabified /s/–obstruent both liaison and enchainment conditions. Finally, there
pairs, which appeared to be identical to the equivalent may be differences between consonant–vowel and vowel–
word-initial sequences. These acoustic factors correlated vowel transitionsthatinhibitrecognitionofthesecondvowel
with performance in a word-spottingtask, suggesting that in the syllable-aligned condition. To test these alternative
listeners can make use of resyllabification information in explanations,we would need a task in which the identifi-
word boundary information. Furthermore, Spinelli et al. cation of the same word-initial information in the target
(in press) have demonstrated a shortening of the liaison words was examined,but in a case in which recognitionof
consonant(15%), as compared with similar nonliaisonse- the relevantwords would fail, denyingaccess to high-level
quences. Consonantsin enchainmentsequenceshad an in- information. Vroomen and de Gelder (1999) were faced
termediate duration.
witha similarissuein theirstudyof resyllabificationeffects
Similar analyses were conducted post hoc on the stim- in Dutch. Theirresponse was toembed thecriticalsegments
uli from our experiments. For the liaison and enchainment in nonsense words, prohibiting lexical involvement in
stimulussets, the durationsof the resyllabified consonants their task (phonememonitoring). This manipulationelim-
were measured by analysis of the waveforms and spectro- inatedthe disadvantagefound for the resyllabifiedstimuli,
grams. The consonantswere measured from the end of the confirming the lexical involvement in the original result.
second and third formants of the preceding vowel until the Similarly, Frauenfelder and Content (1999) demonstrated
beginning of the second and third formants of the follow- that the Mehler et al. (1981)syllabificationeffect in French
ba.lance
bal.con
vs. ) was eliminated when the target
ing vowel. These items were matched(for 29 of the 30 sets) (e.g.,
on the resyllabified consonants(14 /t/, 7 /r/, and 8 /z/), but syllables were embedded in nonwords (e.g.,
pal.nuffe
pa.loune
vs.
). Frauenfelder and Content argued that in the ab-
not on the preceding segment. The consonantshad a mean
duration of 73 msec in the liaison condition and 74 msec sence of a match between speech input and word repre-
in the enchainment condition (no significant difference). sentations,the way in which speech is syllabifiedbecomes
These values are similar to the values extracted from the unimportant. Instead, the task becomes one of simply
Spinelliet al. (in press) stimuli (70 and 75 msec), for which finding a phonemic match to the target sequence.
the same speakerwas used. Our experimenthad no matched
Experiment 3 applies this “delexicalization”methodol-
test stimuli in which the critical consonants were word ini- ogy to thepresentstudy. To ensurethat the precise acoustic
tial, so we selected all items from the filler materials that propertiesof the critical segments were preserved, excised
contained a /t/, /r/, or /z/ in a word-initialposition(19 items: sections of the original recordings from the first two ex-
11/t/, 3 /r/, and 5 /z/). Despitethisset’shavinga smaller pro- periments were used. The beginnings and ends of each
portion of the phoneme with the longest duration (/t/), it recorded phrase were removed, makingthe originalwords
un virtu-
still had a mean duration of 88 msec, which was similar to difficult to recognize. For example, the ends of
ose italien
tuose ita
the comparable condition in Spinelli et al. (82 msec) and
were cut out, leaving
. This manipula-
was significantlyhigher than the means for the two sets of tion may not completely eliminate a lexicalstrategy, but it
t
5
resyllabified consonants [for both analyses, (47) 2.2, will certainly reduce its influence. To further discourage
p ,
.05]. These post hoc analyses, therefore, agreed with lexical processing, the task of the participant was altered
other studies showing a reduction in consonant duration from word monitoring to phoneme sequence monitoring.
for cases of liaison and enchainment. It is plausible that The target fragment in each case was the initial vowel of
these cues are combined with lexical information to im- the target, plus the following consonant (e.g., instead of
italien
it
prove the prediction of the following segment. Moreover,
, the participant would monitor simply for ). This
if these acoustic differences are particularly salient, they task was preferred over monitoring for the vowel alone,
may provide a cue to the upcoming vowel even in the ab- because it reduced task difficulty.
sence of lexical knowledge about the previous word.
A fourth possibility with an even stronger focus on the
EXPERIMENT 3
acoustic information in the signal is that the conditions
vary simply in terms of the intelligibility of the speech.
Method
For example, it is possible that transitions between adja-
Participants. Forty-five paid native speakers of French took part
cent vowels make the second vowel difficult to identify, in
in this experiment. The participants were recruited in Cambridge
which case the effects shown in Experiment 2 would have
to be reevaluated.
through the university and language schools, and none had partici-
pated in either of the first two experiments.

806
GASKELL, SPINELLI, AND MEUNIER
F
5
significant by participants, but not by items [ 1(2,72)
Table 4
Results of Experiment 3 (Fragment Monitoring)
p ,
F
5
p 5
3.7,
.05; 2(2,52) 2.1, .13], with the liaison
condition showing the fewest errors, as well as the fastest
responses.
Measure
Liaison
Enchainment
Syllable Aligned
RT (msec)
Error rate (%)
730
7.1
787
9.6
778
12.5
Discussion
Experiment 3 was intended to provide preliminary evi-
dence for the source of the facilitation effects for resyl-
labified speech in word monitoring in Experiment 2. We
suggested four possiblecauses of these effects, which var-
ied in terms of the type of information required. Of these
four, two options now appear unlikely. If the effect in Ex-
periment 2 was due to a relative difficulty in picking out
target vowels from vowel–vowel transitions,as opposed to
consonant–vowel transitions, the pattern of response times
should have remained unchanged. Although there was
reasonable evidence that the liaison advantageover sylla-
ble aligned remained in Experiment 3 (a 48-msec, or 6%,
reduction in response speed, as compared with 44 msec,
or 9%, in Experiment 2), the difference between the en-
chainment and the aligned conditions vanished entirely
Stimuli and Design. The recordings from Experiment 2 were
used, but with the beginnings and ends cut off to make the use of lex-
ical information more difficult. The article (un) was removed, and a
syllable was cut from either end, ensuring that the second word in the
phrase was affected, plus the target from the word-monitoring ex-
periment. There was always at least a syllable of the target word re-
maining, and often more. The targets for the test sentences were the
initial vowel and consonant of the earlier target word (e.g., ob from
objectif ). In some cases, the target was contained within a single syl-
lable, and in others, the target included the initial consonant of the
next syllable (e.g., am in amateur). In one case (ex in exemplaire),
the target contained two consonants. Given the null effect found in
Frauenfelder and Content (1999), these syllabification differences
were assumed to be unimportant. In all, there were 30 test items and
90 fillers, plus 10 practice items. The fillers were constructed and
excised in the same way as the test sentences, and each target item
was paired with 3 filler items employing the same target, 1 of which
actually contained the target. In this way, 50% of the target items
were present in the sentences (30 test 1 30 fillers). The target se-
quences were recorded using a different native French speaker. As
before, three experimental lists were used in order to counterbalance
conditions and avoid repetition of targets.
Procedure. The participants were first familiarized with the audi-
tory targets by presenting them one by one, accompanied by a writ-
ten list of three words that contained those sequences (e.g., for the tar-
get it, the familiarization words were item, agiter, and marmite). The
words in the list exemplif ied the range of spellings that corre-
sponded to the phonological target, contained the target at different
points in the word, and used different syllabifications. Each target
sequence was presented twice, with participants controlling the pace
of the familiarization.
2
2
9 msec, or 1%, as compared with 36 msec, or 7%, in
(
Experiment 2).
The plausibility that syntactic differences between the
liaison condition and the other two conditions influenced
response speed in Experiment 2 is also further weakened
by the results of Experiment 3, although in this case, it
cannot be ruled out entirely. The excision of the begin-
nings and ends of the phrases eliminatedthe availabilityof
syntacticinformation,yet there remained a marginal effect
of context type, which appeared to be due largely to fast
responses in the liaisoncondition.Nonetheless,this effect
was only marginal, and there remains the slight possibility
that syntactic differences in the compositionof the stimu-
lus phrases had consequences for the prosody of the
phrases and that these prosodic differences remained in
the auditory stimuli for Experiment 3.
The detection task followed. The participants were asked to detect
the target sequence in the shortened phrase. The target was presented
through headphones, followed by a 1,500-msec gap and a visual cue
(a cross on the screen). Next, the phrase fragment was presented au-
ditorily. The participants performed the monitoring task by pressing
a response button if they heard the target. Response latencies were
measured (as in Experiment 2) from the onset of the target vowel.
The results of Experiment 3 undermine the two rela-
tively uninteresting explanations of the resyllabification
advantage shown in Experiment 2. However, we still can-
not fully determine the source of the facilitatory effects of
resyllabification on monitoring for vowel-initial words.
Results
Unsurprisingly, performance in this experiment was We argued that recognition of the word preceding the re-
slower, and there were more errors, than in Experiment 2 syllabified consonant should lead to speeded recognition
(see Table 4). Six participants showed error rates above for enchainment, but not for liaison. Experiment 3 sup-
25%, and their data were excluded from further analysis. ported this explanation,because the preceding words were
One item in the enchainment condition was erroneously
presented with full sentential context and was therefore
excluded as well.
made unrecognizable by excision of their onsets and this
manipulationeliminated the enchainment advantage over
syllable alignment. Equally, the residual (statistically
Despite the removal of the initial and final portions of weaker) advantage that appeared to remain for the liaison
the auditory stimuli, the liaison condition showed a 48- condition could be explained by the presence of acoustic
msec advantage over the syllable-alignedcondition and a markings of liaison that were available in both Experi-
57-msec advantage over the enchainment condition.
ANOVAs on the response time data revealed an effect of
ments 2 and 3 to facilitate responses. Therefore, there may
be two sources of information working, with stronger ef-
target context that was significant by participants,but not fects of lexical knowledge in the case of enchainment and
p , p 5
F
5
F
5
by items[ 1(2,72) 6.3,
.01; 2(2,52) 2.2,
.12]. stronger acoustic cues in the case of liaison. However, an
In the error analysis, the effect of target context was again alternativeexplanationis that our manipulationof the spo-

PERCEPTION OF FRENCH RESYLLABIFICATION
807
ken stimuli has simply made durational (and possibly (Gaskell& Marslen-Wilson,1998;Gow, 2001;Quenéet al.,
lean
other) acoustic cues harder to identify, which has resulted 1998), in which the observation of variation (e.g.,
bacon
leam bacon
in an across-the-board weakening of the resyllabification
becomes
) triggers the expectation of
advantage. Evidence from further experiments will be the context that caused that change, and vice versa. Such
required if these suggestive accounts are to be properly constraint-basedinference is easily learned by a statistical
validated.
learning mechanism, such as a simple recurrent network
(Gaskell, Hare, & Marslen-Wilson, 1995).
GENERAL DISCUSSION
Gaskell and Marslen-Wilson (1998) advanced two po-
tentialmechanisms for the contextualuse of phonological
In this research, we addressed an important test case for variation. A prelexical mechanism relies only on observa-
the syllable as a unit in the perception of romance lan- tion of the variation to predict the viable context. A lexical
guages, such as French. In all four experiments, we found mechanism relies on identification of the word carrying
no evidence for a processing cost of misalignment be- the variant. Gaskell and Marslen-Wilson found that when
geam
tween word and syllable boundaries caused by liaison or carrier sequences were changed into nonwords (e.g.,
bacon
leam bacon
enchainment. The two priming subexperiments showed
instead of
), the contextual effect was
equally strong lexical activation for syllable-aligned and weakenedbut not eliminated.They argued that both lexical
syllable-misalignedwords during the perception of those and prelexicalinformation sources were combinedin order
words. Using word monitoring, we found a clear advan- toresolvetheambiguitycausedbythephonologicalchange.
tage for the misalignedconditionsin the recognitionof the
A similar issue arises in the case of resyllabification. If
next word. Finally, using sequencemonitoringwith the be- durational changes are sufficiently obvious, they can be
ginnings and ends of the spoken stimuli removed, we used to predict the appropriate contextwithoutrecourse to
found no cost for enchainment and weaker evidence for a lexical knowledge. This fits in with the weak liaison ad-
benefit in the case of liaison. Therefore, none of the ex- vantagein Experiment3, in which lexicalinformationwas
periments showed any support for the idea that resyllabi- largely absent. A number of lines of evidence, however,
fication impairs or delays the recognitionof the following suggest that acoustic influences, on their own, cannot un-
word, and the only effects we found were suggestive of a derlie the full pattern of effects found here. First, the
benefit for resyllabification. Here, we will discuss the im- acoustic analyses themselves suggest that consonant du-
plicationsof these findingsfor theories for models of seg- ration alone will not be an infallible indicator of resyllab-
mentation and word recognition. First, however, we will ification. Although there is a clear difference between the
speculate on the possible causes of the facilitative effects mean durationsof resyllabifiedand normal syllable-initial
found for syllable–word boundary alignment with word consonants, the distributions are very much overlapping.
monitoring in Experiment 2 and, more weakly, with se- In our stimuli, 17% of the liaison consonants had a dura-
quence monitoring in Experiment 3.
tion as long as or longer than the mean of the word-initial
It is currently unclear exactly what information sources consonants,and 83% had a longerduration than the short-
underpin the benefits found, but they are likely to include est initial consonant we examined. Unless there are other,
some combinationof lexicaland acousticcues to the pres- more discriminating cues or combinations of cues, this
ence of resyllabified consonants, allowing the perceptual suggests that acoustic information about consonant dura-
system to anticipate a valid phonological context for the tion can only be used probabilistically.
resyllabification. Our stimuli, like others, show a reduc-
Furthermore, although the effects of resyllabification
tion in consonant duration associated with resyllabifica- remained for the liaison condition in Experiment 3, they
tion, and there may be other, more subtle markings of re- were eliminated for the enchainment condition. The pre-
syllabified consonants. Of course, simply identifying an dictive information in the resyllabified enchainment con-
acoustic difference between resyllabified and normal sonantswas not sufficient to generate strong facilitationof
syllable-initial consonants tells us nothing about whether responses to the target sequences when lexical access
this difference is exploited by the perceptual system. But failed. Although the source of the facilitation effects
given the striking sensitivity of the perceptual system to found in these experiments is not fully clear, our best es-
subphonemicinformationin similar domains(e.g., Davis, timate at present is that lexical and acoustic factors are
Marslen-Wilson, & Gaskell, 2002; Hawkins & Nguyen, both influential, with lexical influences critical in the en-
in press), it would be surprising if this informationwas ig- chainment advantage found in Experiment 2.
nored. If listeners do exploitthese durationalcues, a word
Finally, Dejean de la Bâtie and Bradley’s (1995) results
boundarycan be predictedfrom the observationof a short- can be interpreted as a tendencyto compensate for liaison
ened consonant, and a suitable following segment (i.e., a even in the absence of any acoustic cues to liaison. As we
liquidor a vowel) can also be predicted. Both of these fac- mentioned,their study showed that responses were slower
excellenttableau
tors would assist in the recognition of a following vowel- to potentialliaison sequences(e.g.,
) than
vrai tableau
initial word. The situation here is similar to the effects of they were to unambiguous sequences (e.g.,
).
between-word place assimilation in English and Dutch This can be explained as a tendency to treat the initial /t/

808
GASKELL, SPINELLI, AND MEUNIER
tableau
of
as a liaison consonant, increasing activation of ments of the vowel-initialsegments. In all cases, these seg-
“a”-initial words, despitethe absence of any acousticcues ments are identified, and the correct word is recognized.
to liaison.
The onlydifferenceis thattheresyllabifiedconsonantseems
It is worth noting that although the two studies most to provide a small and transient advantage in speed of
closely related to the present research (Dejean de la Bâtie identification of the following vowel.
& Bradley, 1995; Vroomen & de Gelder, 1999) show in-
hibitoryeffects ofresyllabificationprocesses,theirfindings Theories of Segmentation and Word Recognition
are not incompatiblewith ours. As we have mentioned,De-
We began this article by listing possible theoreticalrea-
jean de la Bâtie and Bradley’s results can be interpreted in sons that resyllabification may cause problems for under-
terms of the observation of a potential liaison’s strengthen- standing the following word. Given that we have experi-
ing the hypothesisthat there is a vowel-initialword follow- mental results demonstrating that, at worst, liaison and
ing the consonantin question.The word-monitoringresults enchainmentdo not interfere with recognitionand, at best,
here (and to a lesser extent, the sequence-monitoring re- these processesfacilitaterecognition,we must now reeval-
sults) suggestthat the same process is operatingin the case uate the theories in this new light.
of actual resyllabifications. The Vroomen and de Gelder
The PWC (Norris et al., 1997) states that a potential
(1999) study showed inhibitory effects on the perception segmentationof the speech stream is disfavoredif it leaves
of the resyllabified consonantitself, whereas our study ex- a consonant or consonant sequence stranded between a
amined the effects of resyllabification processes on the known word boundary and a hypothesized boundary. In
following context. The picture from the combined results the case of romance languages, known or likely word
suggests a certain symmetry in the informational content boundaries would include all syllable boundaries. Norris
of resyllabification. The process results in increased am- et al. noted that this constraint could cause problems for
biguityoverthe status of the consonantitself, but once that the perception of liaison in French and suggested that
ambiguity has been resolved (possibly by combining lex- acoustic and lexical information could be used to identify
ical knowledgewith acoustic information), there is a con- liaison consonants and “switch off ” this constraint. The
sequent informational advantage in the predictability of present results are not at odds with the PWC, because li-
the following contextthat licenses the resyllabification. A aison and enchainment are seen as exceptional cases for
similar picture has emerged in the perception of place as- which the constraintdoes not apply. The word-monitoring
similationand the anticipationof its phonologicalcontext data, however, do refine the role of resyllabificationin the
(Gaskell & Marslen-Wilson, 1998; Gow, 2001).
PWC. They suggest that when resyllabified consonants
In summary, although further research is necessary to are identifiedby the system, theresponse is more than sim-
clarify the conditions under which resyllabification in ply a blocking of the standard strategy. The information
French can facilitate speech processing, converging evi- appears to be used positively, to predict the context that li-
dence from a number of sources suggests that the predic- censes the phonological process and to boost (very tran-
tive use of resyllabificationcues relies on some combina- siently) the activation of words that contain the relevant
tion of two types of information—acoustic correlates of context. Given that the perceptual system is required to
resyllabification and lexical information about word identifythese consonantsin order to inhibiterroneous seg-
forms and their boundaries. Although there is some evi- mentations of the speech stream, it seems quite natural
dence that each can operate in the absence of the other, it that the information should also be used beneficially, to
may be that both are used in the normal course of events hone predictions about the upcoming speech.
to assist in the segmentation and recognition of words.
The second theory that might predict inhibitory effects
This gives a boost to the immediate activation of words of resyllabificationis thesyllablehypothesis(Mehleret al.,
that providea viablecontextfor resyllabificationprocesses, 1981). Two different strands of this theory have emerged,
as demonstrated in Experiment 2. However, given the one relating to the basic units of representation in lexical
equally strong priming in all the test conditions in Exper- access and a second relating to word segmentation. Both
iment 1, it is important to bear in mind that this initial ad- these strands would, on the face of it, predict that resyl-
vantage is transient. The monitoring experiments primar- labification should interfere with the recognition of the
access
ilymeasuredthespeedofrecognitionof theinitialsegments following word. The
strand of the syllable hypoth-
of the target words. The results of Experiment 1B showed esis would require complex mappings between syllable
that by halfway through the next word, the advantage for units and lexical entries in order to allow the correct word
the resyllabificationconditionshad disappeared. It seems to be recognized using only syllabic units (e.g., for an en-
that the normal course of lexical activation and competi- chainment sequence, the syllable containing the resyllab-
tion is sufficient to activate the correct lexical items for ified word would need to be part-mapped onto two sepa-
the stimuli in the syllable-aligned condition of Experi- rate words). Alternatively, theaccess version of the syllable
ment 1. Thisis perhaps notsurprising,giventhat in allcon- hypothesiscould allow for smaller units of representation
ditions, there is abundant information available in the to dominate in cases like this (cf. Sebastian-Gallès et al.,
speech signal that can be used to identify the initial seg- 1992), but even then one would expect the reliance on an

PERCEPTION OF FRENCH RESYLLABIFICATION
809
alternative representation to have a cost in terms of pro-
cessing time comparable to the syllablemisalignmentcost
in Mehler et al.’s original research.
It may be more advantageousto turn instead to the seg-
mentationversion of the syllable hypothesis.Given a lack
of preferential evidence for the access version of the syl-
lable hypothesis,plus their own evidencefavoringthe seg-
mentation version, Content, Dumay, and Frauenfelder
(2000;see also Content,Meunier, Kearns, & Frauenfelder,
mentation: Evidence from juncturemisperception. Journal of Memory
& Language, 31, 218-236.
Cutler, A., & Carter, D. M. (1987). The predominance of strong ini-
tial syllables in the English vocabulary. Computer Speech & Lan-
guage, 2, 133-142.
Cutler, A., Demuth, K., & McQueen, J. M. (2002). Universality ver-
sus language-specificity in listening to runningspeech. Psychological
Science, 13, 258-262.
Cutler, A., Mehler,J., Norris, D., & Segui, J. (1986). The syllable’s
differing role in segmentation of French and English. Journalof Mem-
ory & Language, 25, 385-400.
2001) argued for the syllable onset segmentation hypoth- Cutler, A., & Norris, D. (1988). The role of strong syllables in seg-
mentation for lexical access. Journal of Experimental Psychology:
Human Perception & Performance, 14, 113-121.
Davis, M. H., Marslen-Wilson, W. D., & Gaskell, M. G. (2002).
Leading up the lexical garden path: Segmentation and ambiguity in
esis (SOSH) in French. SOSH assumes that lexical access
relies on relatively small units of access, such as features
or phonemes, but that the syllable onset is a privileged
alignmentpointfor a lexicalsearch. Their model takes ad-
vantage of the strong correlation between syllable and
word onsets in French. A simple application of this strat-
egy would again lead to the prediction that resyllabifica-
tion causes problems for recognition, and indeed this
prediction has been confirmed in the case of the resyl-
labification before consonants such as /r/ and /l/ (Dumay
et al., 1999). However, as in the PWC model, if resyllabi-
fied consonants can be identified as such, either through
the applicationof lexical knowledge or the recognition of
acoustic markings, it is possible that a more flexible ver-
sion of SOSH could accommodate our results.
In summary, the lack of a resyllabification cost in the
perception of vowel-initial words and the suggestion of
some transient advantages associated with enchainment
and liaison demand a flexible view of the role of the sylla-
ble in the perception of French in particularand in speech
perception in general. Our results appear incompatible
with the idea that syllable is the basic unit of perception
for French listeners. Instead, they suggest that syllable
onset should be thought of as one of several cues to likely
word onsets. For vowel-initial words, it appears that we
can use other information (e.g., lexical, acoustic, and
spoken word recognition. Journal of Experimental Psychology:
Human Perception & Performance, 28, 218-244.
Dejean de la Bâtie, B., & Bradley, D. C. (1995). Resolving word
boundariesin spoken French: Native and nonnativestrategies. Applied
Psycholinguistics, 16, 59-81.
Dumay, N., Content, A., & Frauenfelder, U. H. (1999). Acoustic-
phonetic cues to word boundary location: Evidence from word spot-
ting. In Proceedings of ICPhS’99 (pp. 281-284). Berkeley, CA: Uni-
versity of California, Linguistics Department.
Dumay, N., Frauenfelder,U. H., & Content, A. (2002). The role of
the syllable in lexical segmentation in French: Word-spotting data.
Brain & Language, 81, 144-161.
Foss, D. J. (1969). Decision processes during sentence comprehension:
Effects of lexical item difficulty and position upon decision times.
Journal of Verbal Learning & Verbal Behavior, 8, 457-462.
Frauenfelder,U. H., & Content, A. (1999). The role of the syllable
in spoken word recognition: Access or segmentation. In Actes des
2ièmes Journées d´Études Linguistiques (pp. 1-8). Nantes, France:
Université de Nantes, Équipe AAI.
Gaskell, M. G., Hare, M., & Marslen-Wilson, W. D. (1995). A con-
nectionistmodel of phonologicalrepresentation in speech perception.
Cognitive Science, 19, 407-439.
Gaskell, M. G., & Marslen-Wilson, W. D. (1998). Mechanisms
of phonological inference in speech perception. Journal of Exper-
imental Psychology: Human Perception & Performance, 24, 380-
396.
Gow, D. W. (2001). Assimilation and anticipation in continuousspoken
word recognition. Journal of Memory & Language, 45, 133-159.
phonological)to compensate for resyllabification and its Grosjean,F. (1996).Gating. Language& CognitiveProcesses, 11, 597-
604.
consequent misalignment of syllable and word bound-
Hawkins, S., & Nguyen, N. (in press). Effects on word recognition of
syllable-onsetcues to syllable-codavoicing.In J. K. Local, R. A. Ogden,
& R. A. M. Temple (Eds.), Papers in LaboratoryPhonologyVI. Cam-
bridge: Cambridge University Press.
Imbs, P. (1971). Trésor de la langue Française: Dictionnaire des fré-
quences [Resource of the French language: Dictionary of frequen-
cies]. Paris: Klincksieck.
aries. It may even be the case that these cues can operate
in the listener’s favor and briefly facilitate the recognition
of words misaligned in this way.
REFERENCES
Bradley, D. C., Sánchez-Casas, R. M., & García-Albea, J. E.
(1993). The status of the syllable in the perception of Spanish and
English. Language & Cognitive Processes, 8, 197-233.
Brent, M. R. (1999).Speech segmentation and word discovery:A com-
putational perspective. Trends in Cognitive Sciences, 3, 294-301.
Cairns, P., Shillcock, R., Chater, N., & Levy, J. (1997). Bootstrap-
ping word boundaries: A bottom-upcorpus based approach to speech
segmentation. Cognitive Psychology, 33, 111-153.
Kilborn, K., & Moss, H. (1996). Word monitoring. Language & Cog-
nitive Processes, 11, 689-694.
McQueen, J. M. (1998). Segmentation of continuous speech using
phonotactics. Journal of Memory & Language, 39, 21-46.
Mehler, J., Dommergues, J. Y., Frauenfelder, U., & Segui, J.
(1981). The syllable’s role in speech segmentation. Journal of Verbal
Learning & Verbal Behavior, 20, 298-305.
Morais, J., Content, A., Cary,L., Mehler,J., & Segui,J. (1989). Syl-
labic segmentation and literacy. Language & Cognitive Processes, 4,
57-67.
Content, A., Dumay, N., & Frauenfelder,U. H. (2000). The role of
syllable structure in lexical segmentation: Helping listeners avoid
mondegreens. In A. Cutler, J. M. McQueen, & R. Zonderon (Eds.), Norris, D., McQueen, J. M., Cutler, A., & Butterfield, S. (1997).
Proceedings of the SWAP conference (pp. 39-42). Nijmegen: Max
Planck Institute for Psycholinguistics.
Content, A., Meunier, C., Kearns, R. K., & Frauenfelder, U. H.
(2001). Sequence detection in pseudowords in French: Where is the
syllable effect? Language& Cognitive Processes, 16, 609-636.
Cutler, A., & Butterfield, S. (1992). Rhythmic cues to speech seg-
The possible-word constraint in the segmentation of continuous
speech. Cognitive Psychology, 34, 191-243.
Pollatsek, A., & Well, A. D. (1995). On the use of counterbalanced
designs in cognitiveresearch: A suggestionfor a better and more pow-
erful analysis. Journal of Experimental Psychology: Learning, Mem-
ory, & Cognition, 21, 785-794.

810
GASKELL, SPINELLI, AND MEUNIER
Quené, H., van Rossum, M., & van Wijck, M. (1998, December). As-
similation and anticipationin word perception. Paper presented at the
Fifth InternationalConference on Spoken LanguageProcessing, Syd-
ney, Australia.
Raaijmakers, J. G. W., Schrijnemakers, J. M. C., & Gremmen, F.
(1999). How to deal with “the language-as-fixed-effect fallacy”:
Common misconceptions and alternative solutions. Journal of Mem-
ory & Language, 41, 416-426.
Vroomen, J., & de Gelder, B. (1999). Lexical access of resyllabified
words: Evidence from phoneme monitoring. Memory & Cognition,
27, 413-421.
Yersin-Besson, C., & Grosjean, F. (1996). L’effet de l’enchainement
sur la reconnaissance des mots dans la parole continue [The effect of
enchainment on the recognitionof words in continuousspeech]. L’an-
née Psychologique, 96, 9-30.
Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical
learning by 8-month-old infants. Science, 274, 1926-1928.
Sebastian-Gallès, N., Dupoux, E., Segui, J., & Mehler, J. (1992).
Contrasting syllabic effects in Catalan and Spanish. Journal of Mem-
ory & Language, 31, 18-32.
Spinelli, E., McQueen, J. M., & Cutler, A. (in press). Processing re-
syllabified words in French. Journal of Memory & Language.
Vroomen, J., & de Gelder, B. (1997). Activation of embedded words
NOTES
1. Content, Meunier, Kearns, and Frauenfelder (2001) question the
generality of Mehler et al.’s (1981) original result, arguing that identifi-
cation of syllable boundaries, rather than syllable units, is critical in the
perception of spoken French. We will discuss their alternative model in
the General Discussion section.
2. Cairns, Shillcock, Chater, and Levy (1997) calculated that a seg-
in spoken word recognition. Journal of Experimental Psychology: mentation strategy based on strong syllables should aid recognition of
Human Perception & Performance, 23, 710-720.
50% of all English words, with a false alarm rate of only 5%.
APPENDIX A
Auditory Stimuli for Experiments 1–3
Enchainment Syllable Aligned
Liaison
Control
vais ori
gulier opti
laise orig
nivers opti
vère acci
seau orig
fini opti
ini acci
un mau
un sin
un lé
un g
un par
un pe
un pre
un fer
un gé
un gra
un f
un pré
un se
un glo
un mau
un ré
un
ginal
miste
un ma
un u
un sé
un g
inal
miste
dent
un oi
un in
un m
inal
miste
dent
un fameux industriel
un adversaire exécrable
un concert officiel
un doute incessant
un puissant explosif
un affreux oreiller
un regard expressif
un notaire extravagant
un mystérieux organisme
un discours ambitieux
un bref entretien
ger acci
dent
rand impu
este impul
ou impu
lsif
sif
un f
lsif
fait huma
tit impré
vu
mier exem
plaire
ète huma
tact impré
çon huma
un gar niste
niste
un po
un con
niste
ceau impré
un mor
vu
vu
plaire
voir exem
tron exem
un de
plaire
un pa
un ne
un cha
vent idéa
cept idéa
un con liste
veu idéa
peau ita
lien
liste
liste
néreux ita
tuose ita
lien
lien
un vir
cieux inter
diose inter
un gran prète
li inter
prète
prète
chien
cieux idéa
un jo
un vr
un ta
un sa
un re
un ba
un qua
aux autri
ase autri
ai autri
chien
bleau idé
al
lon em
pire
un v
chien
nase idéa
un gym
l
l
un curieux animal
un réel effort
cond em
deste em
rose anni
versaire
pire
un mo
un mo
pire
rieux ann
dit inso
lent
pas anni
iversaire
versaire
un sérieux attachement
un cruel empereur
un délicieux apéritif
un dangereux enseignement
un étonnant appétit
un triste avenir
tiste inso
lent
cessaire entraîne
ment
narchiste intellec
tuel
litaire habi
tue
ron inso
un ar
un né
un a
un mi
lent
gulier entraîne
éminent intellec
milier habi
tue
si entraîne
ment
tuel
ment
crivain intellec
un é tuel
veau habi
un cer tue
un fa
un l
ong hive
eul hive
chu hive
un fi r
r
un s
r
yeux infir
bèse infir
ent infir
un cli me
un jo
me
un o
me
un flagrant attrait
nier objec
illant emplo
yé
duisant argen
voir objec
cit objec
tif
golo emplo
yé
un der
tif
un sa
tif
un ré
un ri
un fâcheux incident
un heureux évènement
un passionnantépisode
un précédent omnibus
un stupide ordre
un stagiaire efficace
un urgent appel
un liquide incolore
un arbuste inutile
un cher adhérent
miste emplo
un chi
un br
un sé
un ex
yé
tin
plomate argen
gneron argen
tin
tin
un di
un vi
un mu
cellent ama
naliste ama
un jour
sicien ama
teur
teur
me
lent orgueil
fond enn
ui
ssier appre
teur
me
leux
alant hom
nête hom
un hon
mi hom
me
un g
un de
un ga
un mé
dulte orgueil
scon orgueil
leux
ga enn
ui
un vio
un pro
un gro
leux
un a
un fu
neste enn
ui
gaire appren
gent appren
ti
vion enne
mi
in inter
dit
nti
mi
un vul
ti
un a
un a
un v
chant enne
un mé
lote enne
un pi
az inter
mi
os inter
dit
un gr
un g
dit
Note—The control condition was used only in Experiment 1. Italicized sections mark the excised sequences used in Exper-
iment 3.
(Manuscript received January 9, 2001;
revision accepted for publication March 28, 2002.)