transcribed speech corpus

alignment

Fig. 1. Automatic speech recognizer regarded as a tool to automatically select parts of audio corpus, which deviate from an expected representation. These parts are of potential interest for more in-depth linguistic investigation.
provide a more extensive description of pronunciation variants. Recent progress in speech modelling provides the opportunity of using a speech recognizer to help analyse large acoustic corpora, and this is an important aspect of this contribution. Facing alternative pronunciations, it is not certain that the recognizer will prefer the same option as a human. Nonetheless, a given ASR system will consistently make the same decisions over the entire corpus, and can be parameterized to best t the investigators needs. In addition, the obtained outcomes may be better suited for ASR than the observations of an expert: whether for instance a schwa is deleted 1 is often unclear, since our perception is biased by our understanding. Figure 1 gives an overview of how the ASR tool can be used for linguistic purposes: depending on the conguration, unexpected pronunciations may be located. By exploring virtually unlimited speech corpora, more precise or even new knowledge may be produced. Our belief is that better formalised knowledge of pronunciation variant mechanisms will ultimately be helpful for pronunciation modelling and ASR systems. In this paper, we focus on syllabic structures and their variation in a large corpus of French radio interview speech. The aim of this study is to detect sequential pronunciation variants (i.e. variants with different numbers of phonemes), and to relate them to syllabic restructuring. These types of variants are considered the most problematic in the present state-of-the-art of ASR systems since improper alignment reduces the acoustic model accuracy. Instead of limiting the linguistic representations to word and phoneme levels, as often is the case, a syllable level is introduced to describe sequential variants at an intermediary level between words and phonemes typically used by ASR systems. There are many reasons to consider syllables: syllables can be seen as basic speech production units and they play an important role in speech perception. Syllables also allow for a better overall temporal location, whereas phonemes may be hard to locate precisely in time, both for
In this paper, phoneme deletion means no distinct temporal segment can be isolated for a given phoneme.

"canonically" aligned speech
"differently" aligned speech

linguistic

investigation

describe formalize new knowledge
humans and for machines. Moreover phonotactic constraints condition the occurrences of shorter pronunciations, and the intermediate syllable level allows us to examine observed variants with respect to expected syllabic structures. In the next section, syllabication rules of spoken French are described: we emphasise the effects of schwa and liaison on the syllabic structure. Section 3 briey describes the speech corpus and outlines the general methodology of aligning speech transcripts to spoken syllables via written syllables. Section 4 explains the link between the word level and an intermediary written language-based syllable level. Section 5 focuses on the spoken syllable level, and Section 6 presents results on syllable restructuring using the W-syllable alignments. To compare with what happens in other languages, we will refer to recent studies by (Corbin 2003) and (Su and Basset 1998), which compare British English with Taiwanese Mandarin and French, by applying a common methodology based on a manually segmented corpus of 15 minutes (each language) of spontaneous speech.
2 Syllabication rules in French
In psycholinguistics, syllables are often considered as the information processing units of perceptual mechanisms, for acoustic-phonetic decoding (Pallier 1994). Various experiments (using a fragment detection task or other techniques) demonstrated that it is hard to focus ones attention on precise phonemes independently from syllables, suggesting that the latter are identied rst. Evidence is also provided by the history of writing (syllabaries are older than alphabets), word games, childrens speech and production errors such as slips of the tongue, which exhibit numerous constraints of syllable structure (Fowler et al. 1993). Finally, this unit is necessary for putting forward rules governing stress patterns. Yet, syllabication, that is the segmentation of the spoken string into syllables, differs from one language to another: it depends on the linguistic communities conventions, and a universal phonological theory does not exist (Vogel 1982). In English (the rhythm of which is not syllable-timed but stress-timed), researchers do not even agree on the number of syllables in words such as communism, hour, real (Ladefoged 1975); and ambisyllabic consonants (which could belong to either syllable simultaneously) as in salad are common (Cutler et al. 1986). In French, a consonant cannot constitute a syllable, and each syllable contains one and only one vowel. Since Saussure (Saussure 1915), a hundred years ago, various theories have been proposed to account for the tendency of some consonant sequences to be split. According to the so-called Sonority Sequencing Principle, phonemes may be arrayed along a sonority scale according to their vowel-likeness, roughly corresponding to their aperture or degree of loudness (perceived intensity). Vowels are the most 5

sonorous type of phoneme, followed in turn by glides, liquids, nasals, fricatives and plosives. The Sonority Principle stipulates that phonemes located in the beginning of a syllable must have increasing sonorities and that syllabic edges are placed just before the minimum of sonority. But this contradicts another principle, that of maximum onset, in cases such as costume. The Maximum Onset Principle (MOP) stipulates that the syllabic boundary between two vowels separated by consonants is placed so as to maximise the number of consonants in the onset of the second syllable. These consonants, though, must constitute legal clusters, i.e. clusters which may appear at the beginning of a word in the language (Kahn 1976). According to the Sonority Sequencing Principle, consonant clusters containing an /s/ followed by two or more consonants undergo a syllabic break after the latter (e.g. obstruer /Obs.tKy.e/ in French). The tautosyllabicity of this /s/ with regards to the following consonant is controversial, since a French word may begin with what Italian grammar calls impure s (e.g. sport), without being disyllabic. In French, it is traditionally assumed that, irrespective of the grammatical and orthographical word tokenization, each consonant belongs to the same syllable as the vowel immediately following. In particular, a syllabic break falls before an intervocalic consonant, even though resyllabication is not complete in some cases (Fougeron et al. 2002): in V#CV and VC#V contexts (where # denotes a word boundary), acoustic cues may enable distinctions such as cas lgal (legal case) vs cale gale (equal hold), both pronounced /kalegal/, since in the latter case the schwa is generally dropped in standard French. The schwa, which may or may not be spoken (thus inuencing the number of syllables), is one of the most intricate aspects of French phonology (Verney Pleasants 1956, Martinet 1971, Dauses 1973, Dell 1973, Walter 1976, Lacheret-Dujour and Pan 1994, Durand and Laks 2000). Consider the word amener (to bring): it has two or three spoken syllables (/am.ne/ or /a.m@.ne/) depending on whether the /@/ is realised or not. Even if it enables a phonological opposition between words such as pelage (coat /p@laZ/) vs plage (beach /plaZ/), the schwa vowel is generally optional. Liaison is another complicated phenomenon which is directly linked to the syllabication process. Liaison consists in the realisation of a normally mute nal consonant in the context of a following word which begins with a vowel. For example, the word sequence les les (the islands) pronounced /le/ and /il/ in isolation are pronounced /lezil/ in connected speech, and liaison results in a cross-word syllabication /le.zil/. Only a limited number of consonants are used for liaison: /z/, /t/, /n/, /K/, /p/ in order of frequency of occurrences. Cross-word syllabication makes word boundary recognition and thus lexical access perceptually more difcult. How and when is liaison made? We are here in a ticklish eld (Delattre 1966, Fouch 1969, Lucci 1983, Encrev 1988, Eggs and Mordellet 1990, Lon 1993, 6

Table 1 Syllabication rules for French (left column), used by the LIMSI G2P converter (G RAPHON +, which can also syllabify a phonemic string given in input). The syllabic break is noted by a dot, @ stands for a maintained schwa; V={vowels}, L={liquids}, G={glides} O={obstruents: plosives, fricatives or nasals} and C={any consonant}. C{0;4} means 0 to 4 consonants. For each rule, the right columns show a word example and the effect of the syllabication rule on the example. The word parts for which the rule applies are underlined. Syllabication rules Word Pronouniation Syllabication
@ C{0;4}V VV VCV VCGV VOLV VCCV VOLGV VCCGV VCOLV VCCCV VCOLGV VCCCCV VCCCCGV

@. C{0;4}V V.V V. CV V. CGV V. OLV VC. CV V. OLGV VC. CGV VC. OLV VCC. CV VC. OLGV VCC. CCV VCC. CCGV
refroidi ralise image studio public objet emploi victoire esprit expert altruiste expier exploit
K@frwadi Kealiz imaZ stydjo pyblik ObZE Aplwa viktwaK EspKi EkspEK altK4ist Ekspje Eksplwa
K@. fKwadi Ke. aliz i. maZ sty. djo py. blik Ob. ZE A. plwa vik. twaK Es. pKi Eks. pEK al. tK4ist Eks. pje Eks. plwa
djeuner (lunch) and msieur for monsieur (Sir) may be observed, where the drop of an unstressed vowel leads to a resyllabication, transforming simple CV structures into more complex syllabic units. While these phenomena are well known to linguists and speakers of French (e.g. (Lon 1993), their prevalence in spontaneous speech and consequences for ASR are not clearly established. We study these effects with the aid of a speech recognition system that is used to automatically label a large speech corpus, in order to carry out further linguistic analyses. By aligning the data with acoustic word models which allow for pronunciation variation (e.g. optional schwas and liaisons), the 8
Table 2 Syllable types of standard French using C (consonants) and V (vowels) classes (after (Wioland 1985)). The last column shows resyllabication if a schwa is produced. The total number of syllable types is thus reduced from 14 to 8 (in bold). Syllables Example Pronunciation Syllables with schwa
CV CCV CVC V CCVC CVCC VC CCCV CCVCC CCCVC VCC CVCCC CCCVCC CVCCCC
veau gr masse eau grade test hte strie Brest strate ogre ltre strict dextre
vo gKe mas{@} o gKad{@} tEst{@} At{@} stKi bKEst{@} stKat{@} OgK{@} ltK{@} stKikt{@} dEkstK{@}
CV CCV CV - CV V CCV - CV CVC - CV V - CV CCCV CCVC - CV CCCV - CV V - CCV CVC - CCV CCCVC - CV CVCC - CCV

the alignment and sequential variant detection task. The orthographic transcription is used during alignment (instead of a language model during recognition). Since the pronunciation dictionary can contain multiple entries per word, the decoding space during alignment corresponds to a phone graph including all allowable pronunciations. For a given word, all pronunciations are equiprobable, independently of the pronunciation length (no insertion/deletion penalties). More details about the alignment procedure and its reliability can be found in (Adda-Decker and Lamel 1999), where we have shown that the number of detected variants decreases with the number of contexts in the acoustic models. In order to reduce the possibility that the models already incorporate some of the reductions of interest here, we have chosen to use a small set of 130 context-dependent continuous density hidden Markov models (HMMs) with Gaussian mixtures. Our conguration allows a very large number of sequential variants: all sequential variants are included sytematically without concern as to whether or not they are linguistically relevant. This is part of the methodology: since only variants which are explicited before can be observed we need to overgenerate. In practice, only a small number of the numerous variants are observed, thus validating the approach. The most frequent variants were manually checked to verify whether or not they are consistent with what we know/expect: checking is done by listening to part of the aligned segments and looking at spectrograms. With the help of this methodology, we aim to identify syllabic restructuring due to well-known phenomena such as schwa deletion and liaison in French. This also helps to partially validate the approach: if already known and described phenomena are automatically detected, other unexpected items should be considered with care. A further aim is then to identify less described deletion phenomena concerning vowels (i.e. syllable nuclei), consonants or even whole syllables. Figure 3 shows a generic syllable representation composed of an onset and a rhyme. The onset is optional and, if present, may contain a single consonant or a consonant cluster. The rhyme has a mandatory nucleus which corresponds to a unique vowel in French. The coda, like the onset, is optional and may be composed of one or more consonants. The right part gives the structure of the most frequent syllable: the CV-type syllable with a single consonant onset and a rhyme limited to the vowel nucleus. In the following, we describe common syllable restructuring phenomena, due to schwa elision and to the French liaison phenomenon. Figure 4 illustrates how the presence or absence of a schwa changes the syllable structure within the word amener (to bring): the left pronunciation has three open syllables V.C@.CV; on the right, the schwa syllable (/m@/) is deleted, resulting in a two syllable pronunciation, with a closed rst syllable VC.CV. A similar restructuring of open syllables can also occur across word boundaries: for example, the sequence prs de Paris /pKEd@paKi/ (near Paris) corresponds to the syllable strucure CCV.C@.CV.CV. Schwa dele11

4.2 W-syllable pronunciation dictionary
A canonical W-syllable transcription has been derived from the lexical transcription (see rst two lines of example in Table 7). In order to align the W-syllables with the acoustic signal, a pronunciation dictionary with variants is introduced. As we are mainly interested in reduction phenomena (inducing a smaller number of phonemes than theoretically expected), any shorter phone sequence included in the MLC form is allowed (see Table 6). In addition to these variants an optional schwa is added to each pronunciation, and each syllable may be reduced to a simple schwa.

4.3 Optional W-syllables

In the W-syllable pronunciation dictionary, each entry can be reduced to one phoneme. Beyond these reductions, we want W-syllables to be optional: if a W-syllable has not been uttered, it should be possible to skip it. For example the word-nal /tK@/ syllable of the word orchestre (/OK.kEs.tK@/) may completely disappear in a sequence such as orchestre de chambre (chamber orchestra): Alignments are carried out using a W-syllable graph corresponding to the W-syllable transcription, where every other syllable may become optional as shown in Figure 7. 15
Table 5 Different W-syllable types observed in the corpus with their percentage of occurrence. The partial syllable C is mainly due to syllabication carried out on isolated words (W-syll isol column). The W-syll cont column gives corrected full syllable percentages, where partial syllables are glued to the following syllables. Finally, the W-syll+liaison column corrects liaison syllabication. means that the percentage is less than 0.05. syllable type CV V CCV CVC C VC CCVC CCCV CVCC VCC CCVCC CVCCC CCCVC VCCC CCCVCC Table 6 Excerpt of the W-syllable pronunciation dictionary (the left side corresponds to the Wsyllable and the right part to the optional smaller length pronunciations). W-syll. sa tKi pronunciations sa s a sa@ s@ a@ @ W-syll isol 57.6 14.6 9.8 9.2 4.3 2.6 1.0 0.5 0.3 0.2 W-syll cont 63.2 12.0 10.5 10.3 2.0 1.1 0.5 0.3 0.1 W-syll+liaison 68.2 9.7 10.7 7.9 1.1 0.8 0.5 0.3 0.1

Table 8 Most frequent spoken syllable types in French for the canonical S-syll column (from the syllabied MLC sequence)and the aligned S-syll column (from the syllabied aligned phone sequence), as observed in the corpus. Closed syllables are more frequent in the aligned S-syllables. syllable type CV V CCV CVC VC CCVC CVCC CCCV
%canonical S-syll 67.3 11.8 10.5 7.6 1.1 0.6 0.4 0.4
%aligned S-syll 60.4 12.5 9.2 11.6 1.6 1.4 1.4 0.4
Fig. 9. Resyllabication producing closed CVC syllables due to V vowel deletion.
cation results, we could observe that simple vowels (often the schwa) are sometimes aligned with unclearly uttered syllables (e.g. repetitions of word fragments): such alignments produce V syllables.
6 Analysis of alignment results
Syllabic restructuring can result from vowel and/or consonant deletions. Figure 9 illustrates a typical cross-word resyllabication of two consecutive CV syllables into a CVC syllable (e.g. vous recherchez (you are looking for) /vu.K@.SEK.Se/ [vuK.SEK.Se]) due to schwa deletion. This kind of resyllabication is very common, as indicated by the CVC gures in Table 8, which show a signicant increase of 4% (absolute) from canonical to aligned S-syllables. Different questions may arise during alignment analysis: Do consonants disappear more than vowels? Are some types of phonemes more deletion prone? Since the schwa vowel is known to be highly instable and schwa syllables can be considered as weak syllables, do they disappear signicantly more than other syllables? Do 19
Fig. 10. Spectrograms illustrating /@/ and /e/ vowels and /K@/ syllable deletions. Left: full pronunciation of the word retrouver /K@.tKu.ve/ Middle: schwa deletion entailing resyllabication: on veut retrouver ses problmes /.v.K@.tKu.ve.se.prO.blEm/ [vK.tKu] O Right: /K@/ syllable deletion and /e/ vowel deletion on veut retrouver ses proccupations /.v.K@.tKu.ve.se.pre.O.ky.pa.sj / [v.tKu] [prO] O O
the monosyllabic function words with a schwa behave as other schwa syllables (in particular word-nal schwa syllables)? Do the disappearing syllables more often correspond to function words than parts of polysyllabic words? What is the most deletion-prone position of the syllable: word-initial, word-internal or word-nal? In the following, we investigate vowel, consonant and syllable deletions with respect to the W-syllable representation. In contrast to S-syllables, W-syllables guarantee a straightforward link with the lexical level, potentially providing insight about word modelling problems due to the cross-word syllabication.

6.1 Deleted vowels

The global deletion rate measured for vowels is 15%. This rate drops to 6% if schwas are excluded. High deletion rates are observed for /O/ (10%), /u/ (8%), / / E (7%), /E/ (7%), /i/ (5%), /a/ (5%). While vowel deletion may occur at VV sequences within words, it is more typical at word boundaries. Vowels are also prone to deletion in N (nasals) and L (liquids) contexts. Another consonant context favouring vowel deletions corresponds to C_C, where left and right C phonemes are equal or close: six cents (six hundred) /sis / may be reduced to [ss ], and si ctait (if it were). /sisetE/ to [ssetE]. A A Even if the underlying vowel can be identied, there is no distinct segment in the acoustic signal. In Table 9, some examples of observed vowel deletions are reported. Likewise, the spectrograms in Figures 10 and 11 illustrate that the vowels are indeed missing in the acoustic signal, and that this is not an artifact due to the automatic alignment procedure. 20
Table 9 Examples of vowel deletion within words and across words in different contexts. word extraordinaire mais enn jai t word VV /EkstKaOKdinEK/ /mEAf / E /ZEete/ CVC V [EkstKaKdinEK] [mAf ] E [Zete] CC
left or right C nasal or liquid cinma comment personnel voulait /sinema/ /kOm / A /pEKsOnEl/ /vulE/ [sinma] [km ] A [pEKsnEl] [vlE]
left and right C equal or close il allait vous voulez carrire artistique musicien /ilalE/ /vuvule/ /kaKjEKaKtistik/ /myzisi / E [illE] [vvule] [kaKjEKKtistik] [myzsi ] E
miscellaneous - frequent word sequences cest cest pas je sais pas /sEta/ /sEpa/ /Z@sEpa/ [sta] [spa] [Spa]
Vowel deletion seems to be more common in French than in British English, where stressed vowel deletion is avoided, and in Taiwanese Mandarin, where the vowel bears the tone ((Corbin 2003) and (Su and Basset 1998)). This is even more obvious in the most frequent (function) words, whereas the tendency in British English is to drop the initial or nal consonant in monosyllabic words such as him, but. If the central schwa-vowel deletion seems to be shared by the three languages, all vowels can undergo deletion in French. This is not the case in English and Chinese, where low vowels appear to be more resistant, more preserved than the other vowels. 21
Fig. 11. Spectrograms illustrating /ne/ syllable and /o/ vowel deletions. Left: full pronunciation of the word cinma /si.ne.ma/ Right: /ne/ syllable and /o/ vowel deletions. cinmatographique /si.ne.ma.to.gKa.k/ [si.mat.gKa.k]

6.2 Deleted consonants

The average consonant deletion rate is 13%. The deletion rates in onset and coda position differ signicantly: in the onset position the consonant deletion rate is 11%, whereas the coda consonant deletion rate is close to 30%. Our results are in agreement with previous work by Duez (Duez 2003) and comparable to studies of English (Greenberg and Chang 2000). Mandarin of course is different since consonants can only be syllable-initial (if nal nasals are considered as part of the vowels). In syllable-initial position, the most deleted consonant is the voiced fricative /v/ (20%), occuring in frequent words such as vous (you), avec (with), avez, avait (have, had). Liquids and glides are also often deleted in this position (12% to 17%). Deletion rates are lowest for unvoiced fricatives. Liquids account for more than 35% of deletions, whereas they represent 1/4 of consonants. As examples of liquid deletions, the word lm is often pronounced as [m], and the syllable /pli/ as in compliqu (complicated) is aligned with [pi] in close to 25% of occurrences. The truncation of words such as montre (watch or show) and prendre (to take) resulting from the drop of the nal liquid is a wellknown phenomenon in spoken French. The analysis of our data conrms that, for words in -tre and -dre preceding a consonant, the pronunciations [t] and [d] (rather than [tK] and [dK] respectively) are preferred. After the schwa elision in this context, the liquid falls in 240 occurrences, and is maintained together with the plosive in 170 occurrences. This way, too massive a violation of the three consonant law is avoided see (Durand and Laks 2000). The drop of the liquid also occurs in il/ils (he/they) before a consonant. In this context, roughly 30% of these tokens are aligned with the pronunciation [i] (300 occurrences) rather than [il] (700 occurrences). In the XVIIth century, the pronunciation [ifo] for il faut (it is necessary 22

liquid

deletion

open syllable

Fig. 12. Example il voudrait (hed like). Step 1: Liquid deletion in coda position resulting in i voudrait and an open syllable which favours Step 2.

VC CV CCV

unstressed vowel deletion
Fig. 13. Step 2: unstressed /u/ vowel deletion and syllable restructuring. The mechanism is identical to the schwa restucturing mechanism.

Table 11 Examples of partial W-syllable deletion (left) and of nal CCV W-syllable deletion. The deletion rate is particularly high for the negation n (24.8%). The word-nal CCV syllables (in bold) are prone to deletion or at least to reduction. word n l qu d c, s rank 12, 66 W-syll. n l k d s #del. 306 % 24.9 8.5 8.3 7.9 7.2 carrier word exemple capable tre W-syll. Eg z pl@ A ka pa bl@ E tr@ aligned Eg z A ka pa E#del 210 % 18.8 12.6 6.4
(7%), vous (6%), oui (6%), mais (5%) (he, no, you, yes, but) of VC and CV types, or to word endings in CCV (Table 11 right). To measure the link between syllable deletion and syllable position within the word, the corpus has been partitioned into 4 subsets of words with a given number of syllables: the monosyllabic, disyllabic, trisyllabic and polysyllabic (> 3 syllables) word sets. Table 12 shows W-syllable deletion rates for the 4 subsets. In French, the word-nal syllable, if not a schwa syllable, bears lexical stress, at least in a prosodic phrase nal position (Delattre 1965). In the adopted MLC pronunciation formalism, many words have a nal schwa resulting in a weak nal syllable. In each of the 4 subsets, words have been separated depending on the last vowel being a schwa or not. Monosyllabic words roughly correspond to the most frequent function words, at least for the schwa set. We can observe that the deletion rates are highest for these monosyllabic function words, with the deletion rate of schwa function words nearly twice the rate of other monosyllabic words. Whereas in the latter subset the deletion rates are above 10% for V syllables (cf. Table 10), more complex syllable structures are less deletion prone. The deletion rates of nal schwa syllables (11.7-14.3%) are very close to those of the monosyllabic function words (11.3%). This suggests that the monosyllabic schwa function words behave as other schwa syllables (in particular word-nal schwa syllables). Even if part of the automatically found deletions may be be related to modelling problems, others are clearly due to syllables missing in the acoustic signal. As expected, nal syllables (respectively penultimate syllables for schwa-nal words) are the most resistant, as shown in bold in Table 12. The lowest deletion rates are for the penultimate syllables in schwa-nal words: this position is less affected by cross-word coarticulation than the non-schwa nal syllables. Measured deletion rates are somewhat higher for word-initial syllables (4.4-6.6%) than for word-internal syllables (3.7-4.2%). This result deserves some further investigation. The V syllable structure which is frequent in word-initial position, but only seldom observed word-internally, is the main explanation here. For example consider the trisyllabic word set with the 5.5% deletion rate in initial position. Removing V syllable initial words from the trisyllabic word set, the deletion rate drops to 2.9% (233 syllable deletions in word initial position out of 8013). In 25

Table 12 Percentages of measured Wsyll deletions in mono-, di-, tri- and poly-syllabic words. For each N-syllabic word set, schwa nal and non-schwa nal subsets are considered separately. Deletion rates are measured with respect to the total number of syllables in the specied position. High word initial deletion rates are mainly due to V syllables. More complex syllable structure in initial position are less deletion prone than word-internal syllables. monosyllabic word set non schwa #occur 122,895 #del 7,771 %del 6.3 #occur 28,373 schwa #del 3,206 %del 11.3
disyllabic word set non-schwa nal position initial nal #occur 32,232 32,232 #del 2,%del 6.6 2.5 schwa nal #occur 24,063 24,063 #del 1,063 2,820 %del 4.4 11.7
trisyllabic word set non-schwa nal position initial penultimate nal #occur 11,853 11,853 11,853 #del 273 %del 5.5 4.2 2.3 schwa nal #occur 10,303 10,303 10,303 #del 1,478 %del 5.2 1.5 14.3
polysyllabic word set non-schwa nal position initial internal penultimate nal #occur 5,026 6,010 5,026 5,026 #del %del 5.7 3.7 4.6 2.5 schwa nal #occur 4,926 5,987 4,926 4,926 #del %del 4.9 3.7 1.6 13.2
contrast the deletion rate for the V syllables is 11% (419 out of 3840). For the disyllabic set the 6.6% word-initial deletion rate corresponds to 3% for non-V syllables and 13.5% for V syllables. Typical examples, where initial vowel deletion occurs in disyllabic words are avait, avez (had, have), enn (at last), avec (with), alors (then), and for trisyllabic words aujourdhui (today), coutez (listen) in sequences such as vingt ans aujourdhui, au fond aujourdhui, je leur dis coutez, je lai rencontr coutez javais dix ans. 26
Results tend to show that phone deletions depend on the position in the syllable and that the acoustic realisation is correlated with word-position. In future modelling of context-dependent phones for ASR, more elaborated contexts can be conditioned, not only on neighbouring phones, but also on the position of the phone in the syllable and the position of the syllable in the word.
7 Conclusions and perspectives
Whereas we all can cite examples of more or less severe reduction phenomena in spontaneous speech, the pronunciation variants are only partially known and they need more extensive description. An ASR system has been used as a linguistic tool to investigate large speech corpora of tens of hours of speech, and to quantify pronunciation trends. In this contribution, we described a new methodology for carrying out corpus analysis on a syllable basis with W-syllables (obtained by syllabifying maximum length canonical pronunciations of isolated words) and S-syllables (where syllabication is carried out on the phoneme string without considering word boundaries). The number of canonical W-syllables is limited to about 1500, whereas the number of effectively aligned S-syllables is signicantly larger (several thousands). The use of W-syllables allows us to relate the word level syllables to spoken ones, although French (unlike English) is supposed to ignore word boundaries when syllabifying an utterance. The limited number of W-syllables simplies the description of the observed variation and facilitates generalisation. For the different W- and S-syllables used, we found relatively stable syllable structure distributions, with the CV structure accounting for more than half of the data. Whereas French theoretically admits 14 different syllable structures (using C and V classes), the 6 structures CV, V, CCV, CVC, VC and CCVC syllables account for 99% of the corpus. Open syllables (CV, V, CCV, CCCV) account for about 90% of the W-syllables in the corpus. Closed syllables are more frequent in the aligned Ssyllables (16%), which best correspond to what was actually said (the aligned and syllabied phone sequence). In speech, the increase of closed syllables is due to vowel deletions and syllabic restructuring. Whereas syllable deletions are relatively frequent for S-syllables (15%), which always measure full syllables (the vowel nucleus is mandatory in both canonical and aligned S-syllables), W-syllables have a much lower deletion rate of 6% (the vowel is mandatory only in the canonical form). Deletions mainly occur in cross-boundary positions. Concerning W-syllable deletions, most of them are due to highly predictable function words and word endings or to V syllable word beginnings. Investigating syllable-position dependent phone deletions, we could measure vowel deletion rates of 15/6% when including/excluding the schwa. Wheres a global consonant deletion rate of 13% could be measured independently of the consonant position within the syllable, coda deletions are three times as frequent as onset deletions. These different deletion options allow the transmitted word rate to increase without physically increasing the speech 27

rate. Or equivalently, it allows us to utter more words with fewer phonemes. This study suggests that phone-in-syllable and syllable-in-word contexts might be of interest for acoustic phone modelling. This is an important direction for future developments. The perspectives of this study are diverse: the developed framework helps describe and quantify more or less well known linguistic phenomena on a syllable basis. Generic rules can then be formulated to generate pronunciation variants, even for rarely observed or unobserved words, for which variants cannot be estimated statistically. Plausible rules can address word-initial vowel deletion, backward onset-coda transfer and forward onset-onset transfer if the resulting onset is permissible (e.g. cest impossible /sE.t.pO.si.bl@/ [st.pO.si.bl@]. In future work, E E we intend to rene the present approach and extend the analysis of the alignment results. The syllable-based framework can also serve as a tool for manual transcription checking: omitted syllables point to either linguistic phenomena or simply to transcription errors. Finally, this research may contribute to syllable modelling for word fragments and out-of-vocabulary words.

8 Acknowledgement

The audio corpus used in this study was provided by the Radio Archives of INA (Institut National de lAudiovisuel) in the context of a collaboration with their research and experimentation department.

References

Adda-Decker, M., Lamel, L., 1999. Pronunciation variants across system conguration, language and speaking style. Speech Communication 29 83-98. Boula de Mareil, P., 1997. tude linguistique applique la synthse de la parole partir du texte. PhD thesis, University of Paris XI, Orsay. Boula de Mareil, P., et al., 1998. Evaluation of grapheme-to-phoneme conversion for textto-speech synthesis in French.In: Proc. LREC 1998, Grenada, pp. 641-646. Boula de Mareil, P., Adda-Decker, M., Gendner, V., 2003. Liaisons in French: a corpusbased study using morpho-syntactic information. In: Proc. ICPhS 2003, Barcelona, pp. 1329-1322. Corbin, O., 2003. Phoneme deletion in spontaneous British English. In: Proc.ICPhS 2003, Barcelona, pp. 2813-2816. Cutler, A., Mehler, J., Norris, D. Segui, J., 1986. The Syllables Differing Role in the Segmentation of French and English. Journal of Memory and Language. 25, 385-400.
Dauses, A., 1973. tudes sur le instable dans le franais familier. Niemeyer Verlag. Tbingen. Delattre, P., 1965. Comparing the phonetic features of English, Spanish, German and French. Julius Gross Verlag. Heidelberg. Delattre, P., 1966. Studies in French and comparative phonetics. Mouton & Co., Paris/London/The Hague. Dell, F., 1973. Les rgles et les sons. Hermann. Paris. Duez, D., 2003. Modelling Aspects of Reduction and Assimilation in Spontaneous French Speech, In Proc. IEEE-ISCA Workshop on Sponrtaeous Speech Processing and Recognition, 2003. Tokyo. Durand, J., Laks, B., 2000. Relire les phonologues du franais : Maurice Grammont et la loi des trois consonnes. Langue franaise. 126 29-38. Eggs, E., Mordellet, I., 1990. Phontique et phonologie du franais. Thorie et pratique. Niemeyer Verlag. Tbingen. Encrev, P., 1988. La liaison avec et sans enchanement. Phonologie tridimensionnelle et usages du franais. ditions du Seuil. Paris. Fouch, P., 1969. Trait de pronunciation franaise. Klincksieck. Paris. Fougeron, C., Goldman, J.-P., Frauenfelder, U.H., 2001. Liaison and schwa deletion in French: an effect of lexical frequency and competition. In: Proc. Eurospeech 2001. Aalborg. pp. 639-642. Fougeron, C., Bagou, 0., Stefanuto, M., Frauenfelder, U.H., 2002. la recherche dindices de frontire lexicale dans la resyllabation. In: Proc. JEP, Nancy 2002. pp. 125-128. Fowler, C.A., Treiman, R., Gross, J., 1993. The structure of English syllables and polysyllables. Journal of Memory and Language. 32 115-140. Gauvain, J.-L., Lamel, L., Adda, G., 1999. The LIMSI 1998 HUB-4e transcription system. In: Proc. DARPA Broadcast News Workshop. Herndon 1999, pp. 99-104. Goslin, J., Content, A., Goldman, J.-P., Frauenfelder, U.H., 1999. Human and machine syllabication in French: A comparison. In: Proc. 2nd Journes de Linguistique. Nantes 1999, pp. 75-80. Greenberg, S., Chang, S., 2000. Linguistic dissection of Switchboard-Corpus Automatic Speech Recognition Systems. In: Proc. ISCA-ITRW Workshop on ASR. Paris 2000, pp. 195-202. Greenberg, S., Carvey, H., Hitchcock, L., Chang S., 2002. Beyond the Phoneme: A Juncture-Accent Model of Spoken Language. In: Proc. Human Language Technology Conference (HLT), San Diego 2002. Kahn, D., 1976. Syllable-based generalisations in English phonology. PhD thesis, MIT (published by Garland, New York, 1980).