On the Categorical Nature of the Process Involved in Schwa Elision in French
Audrey Bürki¹², Cécile Fougeron¹, Cédric Gendrot¹
¹Laboratoire de Phonétique et
Phonologie, UMR7018, CNRS-Paris3/Sorbonne Nouvelle, France
²Laboratoire de Psycholinguistique Expérimentale, Université de Genève, Suisse
audrey.buerki@pse.unige.ch; {cecile.fougeron; cedric.gendrot}@univ-paris3.fr
Abstract
This paper examines the nature of the process involved in
optional schwa elision in French. More specifically, it aims at
testing whether this process is gradual or categorical, on the
basis of an analysis of the distribution of the duration of over
4000 schwas extracted from a large corpus of continuous
speech. The distribution observed is bimodal, with absent
schwas (at 0 ms duration) on one side, and realized schwas on
the other side, the two groups being separated by a small gap
in the distribution. Realized schwas present cases of strong
temporal reduction, but this reduction does not show a
continuous pattern toward zero duration, as would be
predicted if schwa elision was the end-point of a gradual
reduction process. Different interpretations of this distribution
are discussed.
Index terms: schwa elision, french, categorical process
1. Introduction
Schwa elision in French is defined as the alternation between
two variants of a same lexical entry containing or not a schwa.
For example, the word ‘fenêtre’ (window) can be equally
pronounced as a schwa-full form
or as a schwa-less
form
Various factors are known to affect the
realization of one or the other variant (speech style, speaker
age and dialect, position in word, etc…), although the choice
for one or the other variant is not fully predictable.
The alternation between schwa-less and schwa-full forms
has been interpreted as either a vowel deletion process, or a
vowel epenthesis process (see [1] for other accounts). Be it a
deletion or an epenthesis, traditional phonological accounts
assume that the process is complete, and results from a rulebased segmental process.
As part of a larger project on the nature of schwa and
schwa elision in French, the present study aims at determining
whether experimental evidence can be found to assess the
categorical nature of this process. Only few acoustic or
articulatory studies have addressed this question. They have
usually focused on the phonetic properties of the consonant
sequences resulting from the elision of schwa, asking whether
some residues attributable to schwa remain in cases where the
vowel is apparently absent. Consonant sequences resulting
from schwa elision are thus compared to similar consonant
sequences not resulting from elision. Results usually agree on
the fact that the former differ from the latter. However, while
these acoustic or articulatory differences are always taken to
reflect remnants of the vowel, their existence is not always
interpreted in the same way.
In [2] for example, the observed acoustic differences
between sequences of the type ‘le bar trouvé’ vs ‘le bas
r’trouvé’ (the apostrophe indicating the elision, the bar found
vs the sock found again) are attributed to the fact that in the
second case the sequence retains its phonological nucleus
despite the loss of its segmental content. A palatographic
study, [3] found articulatory differences when comparing
consonant sequences resulting from schwa elision (e.g. ‘pas
d’rôle’, no role) to underlying onset clusters (e.g. ‘pas drôle’,
not funny) and consonant sequences spanning a word
boundary (e.g. ‘jade rose’, rose jade). They do not take these
results as evidence for the maintenance of the syllable, but
rather interpret them as a preservation of the pre-vocalic
underlying articulatory specification of the consonants,
despite the deletion of schwa. In her formalization of these
results, [4] considers that these characteristics of the
surrounding consonants, inherited from the form with schwa,
are extended to the form where the schwa is categorically
elided following Paradigm Uniformity principles.
While these interpretations ([2, 3, 4]) have different
implications, none of them raises doubts about the fact that
the absence of schwa results from a categorical segmental
process. This is not the case of other studies, which consider
the phonetic differences observed in [2] and [3] as evidence
for a gradual reduction process. According to the analysis
proposed for English schwa elision within the Articulatory
Phonology framework [5], schwa elision in French would be
the consequence of a gradual process of overlap between
gestures. According to this view, in apparent schwa-less
French forms, the vowel is not acoustically or perceptualy
visible, not because of segmental deletion, but because it is
hidden by the overlap of the surrounding consonantal
gestures. So-called schwa elision would thus be the end point
of this gradual reduction mechanism. In that sense, it is
similar to the other coarticulatory or allophonic variations,
explained by the overlap of invariant underlying gestures.
Support for this overlap account in French is found in the
results of [6] who observed some lingual gestures attributable
to schwa in apparent schwa-less forms, and in the results of
[7] who observed some remnants of the rounded articulation
of the elided schwa on the surrounding consonants.
Although these results have received contradictory
explanations or have failed to be replicated (see [1]), their
major shortcoming resides in their assumption that schwa
elision in French is a ‘casual speech’ or ‘fast speech’ process.
Their argument is that in casual or fast speech, overlap is
known to increase, and thus schwa has more chance to be
fully overlapped (thus hidden). However, even though it is
well agreed upon that there are more schwa elisions in casual
and fast speech style in French, schwa-less variants do occur
frequently at slow and formal speech styles. It cannot thus be
reduced to a continuous speech process.
In a recent paper, [8] examined the acoustic properties of
consonant sequences resulting from the elision of schwa in
English. Her results confirm the existence of acoustic residues
attributable to schwa in apparent schwa-less forms. Following
[5], she also argued that these residues are better accounted
for by a change in gestural overlap, rather than segmental
deletion. More interestingly, although she manipulated speech
rate to trigger schwa elision, she observed that some of her
speakers also produced elision at normal rates. This leads her
to interpret elision at fast and at normal rate by the same
process of overlap, a categorical shift toward a more
2. Method
A repertory of 18553 French words containing a schwa
(obligatory or optional, not occurring at word or clitic
boundaries or in compounds) was created through the
compilation of several lexical databases (Lexique [9], Brulex
[10], IlPho [11], Grand Robert dictionary). In order to
observe the realization of a large number of lexical forms
produced by many speakers, we selected the ESTER corpus
containing 24h of radio broadcasted news including
productions of a total of 574 speakers [12]. We decided to
discard schwas appearing before derivational suffixes or
inflectional endings.
Our focus was on schwas that can alternate with ∅. Two
types of schwa are described in French: optional (alternating)
schwas and obligatory schwas (always produced). While some
words contain unambiguous obligatory schwas (e.g. ‘belette’,
weasel), for other it is in practice very difficult to assert if the
schwa is optional or not. A first intuitive categorization of a
subset of the corpus according to whether the word could be
‘heard’ without its schwa, revealed a great disparity in the
judgements of the 3 authors. Thus, in order to select our test
items, we decided to choose a more objective method based
on the analysis of the actual realization of these words in our
speech database. On the basis of the transcription given by the
automatic recognition system developed at IRISA, we
selected the words that were recognized by the system as
being produced in the corpus sometimes in their schwa-less
form, sometimes in their schwa-full form. After elimination of
the production of 22 speakers having a strong foreign accent,
and of 95 occurrences showing a clear recognition error or
having a bad acoustic quality, 4320 occurrences
(corresponding to 203 different words) were selected for
analysis.
A manual examination of the output of automatic
recognition system was undertaken for all the sound files in
order to correct schwa detection errors (wrong omission /
insertion). Since the shortest segment duration assigned by
the system is 30 ms, and since the automatic alignments are
far from being reliable with respect to the precision required
in this study, duration measurements could not depend on this
automatic segmentation. Automatic segmentations /
alignments were thus corrected manually, according to the
following criteria. We considered a schwa as being realized if
it presented a voiced portion and a formant structure. Fully
devoiced vowels were thus not considered as schwas. The
onset and offset of the vowels were determined based on the
signal periodicity and on apparition / disappearance of the
second formant, as illustrated in the segmentation shown in
Figure 3. These criteria for segmentation of the vowel
intervals were the best compromise with regard to the wide
variety of consonantal contexts. As can be seen in the
example of ‘semaine’ in Figure 3, the duration of the segment
corresponding to our definition is much shorter than if we had
included voiceless portions of the vowel. The consequences
of this type of segmentation are discussed in section 4. The
duration of the schwa intervals was then automatically
computed with Praat [13].
In order to avoid potential confounds due to factors
known to affect the frequency of schwa elision, the words
present in the corpus were coded according to both their
syllabic position in the word, and their location relative to a
morpheme boundary. Schwas in word-initial syllables are
known to be less frequently elided than non-initial schwa,
while morpheme-internal schwas are more consensually
considered as underlying lexical schwas compared to those
located at morpheme boundaries [see[[14]].
3. Results
3.1. General distribution of duration
The distribution of the duration of the 4185 optional schwas
of our corpus is illustrated in Figure 1. This distribution is
divided into two sub-groups. It excludes 4% of the
occurrences, whose duration was over 2 standard-deviations
away from the mean.
1400
Nb of occurrences
overlapping coordination pattern between gestures. Hence,
vowel elision, although explained by overlap, would not only
result from a gradual reduction process, but could be the
consequence of a categorical change in the specification of
the coordination of the gestures. According to Articulatory
Phonology principles, this specification is assumed to be
assigned by the phonological system. While this hypothesis
has to be further elaborated to be applied to schwa alternation
in French, it reconciles the views that elision may be due to
overlap and that this process would be categorical.
This review of the literature shows that determining
whether schwa elision is a gradual or categorical process in
French is not straightforward. In this study, our goal is to
question the nature of the schwa-zero alternating process, but
our focus is not on the potential acoustic or articulatory
residues of schwa in apparent schwa-less forms. Our approach
is different. Rather than focussing on cases where schwa is
acoustically and perceptively absent, as done in the studies
presented above, we concentrate our attention on the different
realizations of words containing a schwa (in their schwa-less
and schwa-full forms). This is done by observing the
production in continuous speech of a large variety of words
uttered by a large number of different speakers. Our
hypothesis is that variations in the realization of these words
can inform us about the nature of the process. More
specifically, we test whether the realizations of this vowel
follow a continuous pattern of reduction ranging from a full
vowel realization to a fully absent vowel (i.e. the end point of
a reduction continuum), or whether two categories emerge:
cases with full vowels on one side, and with completely
deleted vowels on the other side. Support for a categorical
process would be found in a bimodal distribution with a gap
in the distribution between the 2 groups. Support for a
gradual process would appear in a continuous distribution
(without a gap).
1200
1000
800
600
400
200
0
-10
0
10
20
30
40
50
60
70
80
90
Schwa duration
Figure 1 : Distribution of the schwa durations
On one side, at 0 ms, 1215 occurrences (29%) of the
words containing an optional schwa were produced with a
complete absence of voicing and formant structure. According
to our criteria, these forms are considered as schwa-less. As
an illustration of these cases, a schwa-less occurrence of the
word ‘sera’ (will be) is presented in figure 2. On the other
side, between 16 and 86 ms, 71% of the occurrences present a
vocalic portion corresponding to a schwa. Confirming the
literature, these 2970 realized schwas in our corpus are
relatively short vowels, with a mean duration of 49 ms ( =
15). The distribution of the duration of these schwa-full forms
is unimodal with close to zero skewness (0.13) and kurtosis (0.52) coefficients. Figure 3 shows two examples of schwa-full
occurrences of the words ‘sera’ (will be) and ‘semaine’
(week).
More interestingly, with regard to our acoustic criteria, no
schwas were observed with a duration of less than 16 ms, thus
a gap is observed in the distribution beween 0 and 16 ms.
This gap in the duration distribution between the 2 subgroups is in favour of a categorical process. It has to be noted
that the outliers we removed on the left-hand side of the
distribution represent only 0.7 % of the schwa-full forms
distribution. This removal cannot therefore be responsible for
the gap we observe.
It also has to be mentionned that the distribution is similar
when the words having a schwa in an /R/ context (words
difficult to segment) are excluded previously from the
analysis.
Figure 2: Schwa-less occurrences of the word ‘sera’, will be
(note that the /R/ is devoiced) and ‘appelé’, called.
3.2. Distribution of the schwa duration according to
its position in word and in morpheme
Out of our 4185 occurrences, 3445 words have an optional
schwa in the word-initial syllable (e.g. ‘regarde’ look,
‘devient’ become). For these words, the distribution is similar
to the one presented above. Two groups emerge with a gap in
the duration distribution between 0 and 16 ms. The mean
schwa duration of the 2815 schwa-full forms is 50 ms ( =
15), with close to zero skewness (0.14) and kurtosis (-0.51)
coefficients. This distribution is therefore close to normality
while right-skewed.
The remaining 740 words have a non-initial schwa, placed
either in the second or third syllable of the word. Their
distribution is similar to that of initial schwas with 2 separate
sub-groups and a gap in the distribution between 0 and 17 ms.
According to what is known in the literature [e.g.[14]], these
non-initial schwas have a higher frequency of elision than
initial ones. Indeed, only 155 occurrences (21%) show
acoustic vocalic information. The schwa duration of these
schwa-full forms (43 ms, = 17) is shorter than that of initial
schwas (t(2968) = 5.626, p<0.0001), the coefficients of
skewness and kurtosis are respectively of 0.32 and -0.84.
A separate analysis was also done according to the
position of the schwa relative to morphological boundaries.
Schwas were coded as being either morpheme-internal (3218
occurrences, e.g. ‘cheval’, horse) or at a morpheme boundary
(938 occurrences, e.g. ‘retrouver’, found again), according to
a rather conservative morphological analysis. Three words
(‘devenir’, ‘remarque’ and ‘remarquer’, corresponding to 29
occurrences) for which categorisation was uncertain were
excluded. The distributions of the duration of both
morpheme-internal and non-internal schwas are similar to the
ones presented above: schwa-less forms on one side (31% of
morpheme internal schwas; 24% of non-internal schwas) and
schwa-full forms on the other side, the two subgroups being
separated by a gap in the distribution between 0 and 16 ms.
Taken together, these results show that the typical bimodal distribution of the duration of the schwas in our corpus
is not dependent on the position of the schwa in the word or
of the morpheme boundary. All the types of schwas analyzed
are distributed over two separate subgroups: schwa-full forms
with visible vocalic information and schwa-less forms defined
by a lack of voicing and formant structure.
Figure 3: Schwa-full occurrences of the words ‘sera’ (will
be) and ‘semaine’ (week). Note that the devoiced portion of
the vowel in ‘semaine’ is not included in our segmentation,
according to our acoustic criteria.
4. Discussion and conclusion
Optional schwa elision in French is defined as an alternation
between the schwa-less and the schwa-full form of a word. In
our corpus of radio broadcasted speech, we indeed observed
that 29% of the occurrences of the words are realized with no
voicing and formant structure corresponding to a schwa
vowel. This process of schwa elision is considered in
traditional phonological accounts as a categorical process of
segmental deletion. In this study of the production of 4185
optional schwas, we have shown that there is a gap in the
duration distribution of these schwas. Thus, two groups
emerge with forms realized without a vocalic portion on one
side, and forms realized with a vocalic portion of some
duration on the other side.
Recall that according to our hypotheses, a continuous
distribution, without a gap between schwa-full and schwa-less
forms would have supported the view of a gradual reduction
process, in which the schwa-less forms is the end-point. Our
results, showing a bimodal distribution separated by a gap are
on the contrary in support of the categorical nature of the
process.
However, our results do not allow us to be assertive on
this interpretation. Indeed, a categorical account is challenged
by the fact that the gap we observe between the two groups is
rather small. No schwas occur with a duration lesser than 16
ms. It corresponds roughly to 2 periods for a fundamental
frequency of 125Hz. Even though we have been very strict on
our segmentation criteria, we cannot exclude a measurement
error rate of 1 or 2 periods.
Along the same line, it could be argued that the gap we
observe in the distribution is a by-product of the criteria we
chose to define the acoustical interval corresponding to the
vowel. While voicing and formant structure are the most
salient cues defining a vowel (and the clearest attributes to
consider when dealing with such a large corpus), we cannot
ignore the fact that vowels can reduce greatly in running
speech and can surface without voicing or a clear formant
structure. As stated earlier devoiced vowels have not been
considered here as schwa-full realizations. Nonetheless, we
are reluctant to consider that modifying our segmentation
criteria to include such realizations would have change the
distribution to support a continuous gradual process. The
spectrogram presented in figure 4 illustrates such a case. The
change in intensity between the end of the [v] and the [n]
could be attributed to a devoiced vowel. The portion of noise
with increased energy is around 25ms. Hence, if this
realization was categorized as schwa-full, it would not have
inflated our distribution in the gap region. Similarly, if our
shortest vowels, such as the one of ‘semaine’ in figure 3, had
been segmented to include some potential devoiced or
formant less portions, their duration would have been longer.
This would have changed the shape of our distribution by
increasing the gap between the 2 groups, but not by
decreasing it.
Figure 4 :Realization of the word ‘avenir’ (future)
An alternative explanation to consider that gives a full
account of our data, is that the distribution we observed
results from both a categorical and a gradual process. On the
one hand, a phonological categorical process would account
for the leftmost mode of the distribution. This process
corresponds to the traditional optional schwa elision
phenomena typical of French, in which the lexical variant
produced is schwa-less. Whether this form results from a
segmental deletion at one representation level or whether it
corresponds to a stored lexical variant is beyond the scope of
this paper. On the other hand, a gradual phonetic process of
reduction would account for the gradual temporal reduction
we observe within the group of the schwa-full forms. As do
other segments, schwa is subject to reduction in continuous
speech. We are not addressing here the question of whether
this reduction is due to a mechanism of overlap, gestural
reduction or target undershoot.
This dual interpretation has to be confirmed by the
thorough analysis of the 1215 occurrences categorized as
being schwa-less in our study. Of great interest will be the
distribution of schwa-less forms and the potentially reduced
schwa-full forms of specific lexical items and particular
speaker productions. Further analysis is also needed to verify
whether the dispersion in the realization of the schwa vowels
mirrors that of other French vowels in the corpus.
The authors wish to thank G. Gravier, R. Ridouane and U.
Frauenfelder for their help and suggestions.
5. References
[1] Côté, M.H. & Morrison, G. “The nature of the
schwa/zero alternation in French clitics: experimental
and non-experimental evidence”, Journal of French
Language Studies, to appear.
[2] Rialland, A. “Schwa et Syllabes en Français”, in L.
Wetzels & E. Sezer (Eds), Studies in Compensatory
Lenghtening, Foris Publications, Dordrecht, 1986.
[3] Fougeron C. & Steriade D. “Au Delà de la Syllabe: le
Rôle des Informations Articulatoires Stockées dans le
Lexique pour l'Analyse de la Chute de Schwa”, JEL,
Nantes, France, 122-127, 1999.
[4] Steriade, D. “Paradigm Uniformity and the PhoneticsPhonology Boundary”, In M. Broe & J. Pierrehumbert
(Eds), Papers in LabPhon5, CUP, Cambridge, 1999.
[5] Browman, C. & Goldstein, L. “Targetless” Schwa: an
Articulatory Analysis”, In G. Doherty & R. Ladd (Eds),
Papers in LabPhonII, CUP, Cambridge, 1992.
[6] Smorodinsky, I. “The phonology and phonetics of
schwas in parisian French: an articulatory analysis”,
Texas Linguistic Forum, Vol. 41, 1998, p. 161-171.
[7] Barnes, J. & Kavitskaya, D. “Phonetic analogy and
schwa deletion in French”, Berkeley Linguistics Society,
Vol. 28, 2002, p. 39-50.
[8] Davidson, L. “Schwa elision in fast speech: segmental
deletion or gestural overlap?”, Phonetica, Vol. 63, 2006,
p.79-112.
[9] New B., Pallier C., Ferrand L. & Matos R. “Une base de
données lexicales du français contemporain sur internet :
LEXIQUE”, L'Année Psychol., Vol. 101, 2001, p. 447462. http://www.lexique.org
[10] Content, A., Mousty, P. & Radeau, M. “Brulex, une base
de données lexicales informatisée pour le français écrit et
parlé”. L’Année Psychol., 90, 1990, p. 551-566.
[11] Boula de Mareüil P., Yvon F., d'Alessandro C., Auberge
V., Vaissière J. & Amelot A. “A French Phonetic
Lexicon with variants for Speech and Language
Processing”, 2nd Intl Conf. on Language Resources and
Evaluation (LREC), 273-276, 2000.
[12] Galliano, S., Geoffrois, E. Mostefa, D., Choukri, K.,
Bonastre, J.-F. & Gravier, G. “ESTER Phase II
Evaluation Campaign for the Rich Transcription of
French Broadcast News”, Proc. of InterSpeech 9 ,
Lisboa, Portugal, 2005.
[13] Boersma, P. & Weenink D. “Praat: doing phonetics by
computer” (Version 4.3.14), 2005. http://www.praat.org/
[14] Côté, M.-H., “Consonant Cluster Phonotactics: a
Perceptual Approach”, Ph.D Dissertation, MIT, 2000.