🗣️ 04. Connected speech: enchaînement, liaison, elision, the mute e and assimilation

French words change shape at their joints. A final consonant that is silent when the word stands alone comes back before a vowel (liaison); a final vowel disappears (elision) and the spelling marks it with an apostrophe; a consonant that is pronounced at the end of one word is carried over to the next syllable (enchaînement); the vowel written e comes and goes (schwa deletion); neighbouring consonants share their voicing (assimilation). The written word keeps the shape it has in isolation, so every one of these processes happens in the sound and not on the page, except elision and contraction, which the spelling records. This chapter describes them in that order of visibility. The sounds themselves are in 01. Consonants and 02. Vowels, the syllable and the rhythmic group in 03. The syllable, stress, rhythm and intonation, the letters in 05. The writing system.

1. Why it matters for a text model

Nearly everything in this chapter is invisible in written text, and that is the first fact about it: a system that reads French spelling never sees a liaison consonant, never sees a dropped schwa, never sees a voicing change. Four things still reach the page.

1. Elision and contraction are written. l'homme, qu'il, jusqu'à, au, du, auquel are the spelling of a sound process, and they are the main tokenisation problem of French (§5, §6, 07. Punctuation, typography, capitals, abbreviations and numbers).
2. Liaison decides which words are audible as singular or plural, masculine or feminine. les amis differs from le ami only by /z/; for the spoken language the determiner is often the only carrier of number (§3.7).
3. The h-initial words are a closed lexical class that is learned word by word and is visible only through elision (l'homme against le héros) (§4).
4. Homophony. Because the liaison consonant and the schwa are not on the page, much of the "ambiguity" of French speech is ambiguity that the spelling has already resolved (les hauts /le o/ against les eaux /le zo/; 05. The writing system §6).

The authorities for the description are the public-domain manuals of pronunciation: Passy 1906 (passy-1906), Nyrop 1902 (nyrop-1902) and above all Martinon 1913 (martinon-1913), which devotes a long chapter to liaison. They describe the Paris pronunciation of the years around 1900, in which liaison was more widely made than it is now. The modern evidence is from corpora: the PFC corpus (Phonologie du français contemporain), analysed for liaison by Bosch & Greefhorst (bosch-2016) and for the schwa by Bayles et al. (bayles-2016), Liégeois (liegeois-2016) and Storme (storme-2021).

2. Enchaînement

Enchaînement is the resyllabification of a consonant that is always pronounced at the end of a word: it is attached to the following syllable when that syllable has no onset of its own. French syllables prefer to begin with a consonant (03. The syllable, stress, rhythm and intonation §1), so the boundary of the word is not the boundary of the syllable.

Enchaînement is not optional in the way liaison is: it occurs with every pronounced final consonant, in every register, and the only way to prevent it is a pause or a glottal stop, which is the usual device before an emphatic word, a name or an h aspiré (le | héros) (Passy 1906, passy-1906; Nyrop 1902, nyrop-1902). The consequence is that in running speech the listener cannot rely on word boundaries to find words: l'amour /la.muʁ/ and la mouche /la.muʃ/ share their first two syllables, and un invalide and un nain valide are the same string [œ̃.nɛ̃.va.lid].

Enchaînement applies to the voiced fricative, nasal and liquid finals as much as to stops, and it applies across the "silent" final e as well: la petite amie /la.pə.ti.ta.mi/ takes the /t/ of petite, which is pronounced, while the e is not. Martinon (martinon-1913) points out that this is not liaison in the strict sense: the consonant is not altered and not added; "there is liaison properly so called only for consonants that are normally not pronounced". The terminological line is the same in modern descriptions: enchaînement is resyllabification of a stable consonant, liaison is the appearance of a latent one.

3. Liaison

3.1 The latent consonant

A liaison consonant is a consonant that is written at the end of a word, silent when the word stands alone or before a consonant, and pronounced before a following word that begins with a vowel (or a mute h, or a glide in a few words). The consonant is latent (consonne latente): it belongs to the word and appears only in context. It is attached to the following syllable as an onset: les amis /le.za.mi/, not /lez.a.mi/.

Only a few consonants occur, and the spelling letter does not always match the sound:

* Counts are those of Bosch & Greefhorst 2016 (bosch-2016): in the 291 208 words of conversation they extracted from the PFC corpus (205 speakers, 18 départements), they found 5 444 realised liaisons; /z/, /n/ and /t/ together make 99.5% of them.

Four points about the alternations.

• Voicing change. The letters s, x, z give /z/ (not /s/), and d gives /t/ (not /d/): grand arbre /ɡʁɑ̃.taʁbʁ/, un grand homme, quand il vient /kɑ̃.til/. Martinon explains the d → /t/ historically (the final d of Old French was devoiced when final, so the liaison form is the old final) and the same for the gutturals: when the final consonants were still pronounced, "the gutturals always sounded /k/, the easier one", which is why sang and long link with /k/. He draws the general rule that the stops (muettes) link only with the voiceless sound /p, k, t/, while the fricatives (spirantes) link only with the voiced sound /v, z/ (martinon-1913). The one liaison with /v/ is neuf (neuf ans /nœ.vɑ̃/, neuf heures), against neuf in neuf francs.
• Denasalisation. A nasal vowel before a liaison /n/ is kept in un ami /œ̃.na.mi/, on a /ɔ̃.na/, en avoir /ɑ̃.na.vwaʁ/, bien aimable /bjɛ̃.nɛ.mabl/ (the usual case), but it is oral in bon ami, ancien élève, certain avis in careful speech: [bɔ.na.mi], [ɑ̃.sjɛ.ne.lɛv]. The oral variant reflects the feminine stem (bonne, ancienne, certaine): the liaison consonant of the masculine is the consonant of the feminine (§3.7).
• Opening of the vowel. The final -er of premier, dernier, léger and the infinitives is closed /e/ in isolation and opened to /ɛ/ before the liaison: premier étage /pʁə.mjɛ.ʁe.taʒ/ (Martinon: "the e half opens"). The liaison of the infinitive in -er (aller à Paris /a.le.ʁa/) is now almost confined to verse and oratory (martinon-1913).
• The linked consonant is not always the one that is spelled. There is no liaison /d/ at all (a final d links as /t/), and a final s, x or z links as /z/ even where the stable pronunciation of the same word has /s/ (six, dix: /sis, dis/ alone, /siz, diz/ before a vowel).

3.2 Obligatory liaison

Liaison is obligatory (liaison obligatoire) inside a close grammatical unit whose parts cannot be separated, and where its omission is heard as a mistake, not as a style. The classical contexts (Martinon 1913; the repeated textbook list, which Bosch & Greefhorst take over from Grevisse & Goosse 2011) are:

Three things are to be noticed.

1. Cohesion decides. The more tightly two words are joined, the more the liaison is obligatory. A determiner and its noun are one unit; a clitic pronoun and its verb are one unit; a preposition and its object are very close; a verb and its complement are not. The general principle in Martinon is that "there can be liaison only between words linked in sense, sometimes very closely" (martinon-1913). Syntactic cohesion is the one factor that all later work keeps: the example-based model of Bosch & Greefhorst reaches its best result with a small set of lexical and syntactic features, and adding sociolinguistic features (age, sex, education, job, interview or conversation) lowered both precision and recall (bosch-2016).
2. Frequency strengthens the liaison. Martinon notes for the dental of verb forms that "the more the forms are used, the more the liaison is necessary", and that il est and ils ont without liaison are "certainly incorrect, above all when they are auxiliaries" (martinon-1913). The speakers of the lower classes of his time even made the liaison "where there is none", as in va-t-en, va-t-il (a /t/ inserted after the imperative va, by analogy with a-t-il). The forms va-t-en and a-t-il are now the standard spellings of these liaisons by analogy.
3. Obligatory does not mean always. In their 291 208-word extract Bosch & Greefhorst count 6 971 places where Grevisse & Goosse (2011) prescribe an obligatory liaison; only 4 220 of them (60.6%) show one in the speech of the PFC speakers. For the 856 contexts where the grammar prescribes an optional liaison, 288 (33.6%) are realised. Their rule-based baseline, which predicts a liaison wherever the grammar calls it obligatory or optional, reaches an overall precision of .58 at a recall of .83 (F .68), and for /t/ only .32 at .59; the authors conclude that the rules, which describe the formal language, over-predict liaison in the spontaneous speech of the corpus (bosch-2016, tables 5–6). The "obligatory" class is a property of the formal register and of reading aloud.

3.3 Forbidden liaison

Liaison is forbidden (liaison interdite) where making it would be taken as an error or would change the sense:

• after et: et alors /e a.lɔʁ/, un homme et une femme; the t of et is never pronounced, and Martinon notes that the sound of et is "distinguished from est precisely by the liaison, since the one always links and the other never does, and this since the sixteenth century at least".
• before an h aspiré: les héros /le.e.ʁo/, en haut /ɑ̃ o/, un hibou, des haricots (§4);
• before onze, oui, and foreign names that begin with a glide-like vowel: les onze enfants, il dit oui, les yachts, des uhlans, des yaourts (the y acts as a semivowel/consonant). Martinon adds that one cannot link with a vowel-initial word cited as a word: des "ah!";
• between a plural noun subject and its verb: les enfants arrivent /le zɑ̃.fɑ̃ a.ʁiv/ (no /z/ after enfants), les élèves étudient; after a singular noun: un enfant intelligent /œ̃.nɑ̃.fɑ̃ ɛ̃.tɛ.li.ʒɑ̃/ (no /t/); un client important; after a proper name: Paris est grand;
• after the -s of a noun when no close tie exists: des gens | autour de moi; between the parts of a coordination: les bras | et les jambes;
• singular nouns ending in a silent consonant: Martinon states that nouns "in principle no longer link, except in verse", and that even in verse a d hardly does (un nœud | assorti, blond | ardent) (martinon-1913);
• a liaison /p/ or /k/ across the labial and velar stops: the plain words coup, camp, sang, rang do not link in everyday speech (un coup à boire), with a handful of exceptions (beaucoup à faire, trop à faire, sang et eau) (martinon-1913).

As a rule of thumb, a liaison is blocked by a boundary with a syntactic break, even when no pause is made, and before a word that is cited or emphasised. It is not made across a punctuation mark in reading, though Martinon allows it in verse.

3.4 Optional liaison and register

Between the obligatory and the forbidden there is a large middle zone where the liaison is optional and is a marker of register. Martinon, writing in 1913, already describes it in terms that fit the corpora a century later: "naturally, in view of the number of liaisons, it is here above all that personal taste intervenes, and many liaisons that are necessary in verse are optional in ordinary speech, where hiatus is frequent". The groups concerned are:

• verb + complement: ils sont arrivés, il est ici, elle était en retard, nous allons à Paris, ils mangent un morceau. The third persons of être and avoir and the auxiliaries are the most often linked (il est ici, ils ont eu), the lexical verbs the least (ils mangent | un morceau);
• plural noun + adjective, or noun + noun: des enfants intelligents, des gens aimables, des idées originales;
• polysyllabic prepositions and adverbs: pendant un an, devant elle, toujours, assez intéressant, beaucoup aimé;
• pas and plus, trop, très, si, quand followed by a vowel-initial word: pas assez, trop heureux, quand on veut.

The liaison is a stylistic variable. Its rate rises with formality (reading aloud, public speech, broadcasting, the stage), with age and with social status, and falls in conversation. The Bosch & Greefhorst data show it clearly in aggregate: 1.9% of all word junctures in the free-conversation extract carry a liaison, while the formally prescribed contexts show 60.6% (obligatory) and 33.6% (optional) (bosch-2016). The authors' conclusion is that optional liaison is predicted better by memory of specific examples in context (the lexical identity of the two words and the words around them) than by general rules, which is the exemplar view of Bybee 2001 that they test. The model that works best (the memory-based learner TiMBL on five lexical and syntactic features) gets F = .86 for /z/ and for /n/, F = .55 for /t/, the most variable of the three, and .82 overall (precision .80, recall .85). The auxiliary est shows what "optional" means in numbers: of the 230 cases in their data where est is followed by a noun, adjective or participle, 80 (34.8%) have the liaison, and the model predicts it in 79 (34.3%), almost the same rate (bosch-2016).

The practical conclusion for a text model is that a liaison is a property of a word pair in a register, not of a spelling pattern. Liaison cannot be read off the spelling alone.

3.5 Words with several behaviours

A number of very frequent words have different liaisons in different uses:

Plus is the textbook case of one spelling and several behaviours: /ply/ before a consonant in the negative or comparative (il n'en a plus, plus grand), /plys/ when it means "in addition" or ends a clause (il en veut encore plus), and /ply.z/ with a liaison when it is the comparative before a vowel (plus intéressant). Six and dix are /si, di/ before a consonant, /sis, dis/ at the end of a clause and /siz, diz/ before a vowel; huit is /ɥit/ when final and links with /t/. Martinon (martinon-1913) separately notes that the t of the 3rd-person verb forms is linked more often the more frequent the form is.

3.6 Wrong liaisons: pataquès and cuir

A pataquès is a liaison with the wrong consonant, or a liaison where there is none to make. The word is the standard joke made out of the error it names: Ce n'est pas-t-à moi /sə nɛ pa ta mwa/ for Ce n'est pas à moi, ce n'est point-z-à moi. Martinon, who gives this origin, distinguishes two lesser kinds: the cuir is the addition of a /t/ (il va-t-en ville), the velours the addition of a /z/ (j'ai z-été, quatre-z-yeux). He adds that the velours is softer than the cuir and that "the cuir itself had the pretension of making pronunciation smoother, perhaps as leather softens the razor"; and he notes that "in former times j'ai zété and il leur za dit were admitted by the most distinguished persons", which was no more extraordinary than a-il or aime-il said /ati/ or /ɛmti/ in the sixteenth century before the t was introduced in spelling (martinon-1913). The history is the point: a-t-il and va-t-en are cuirs that became correct.

The modern sociolinguistic term is hypercorrect liaison: a speaker who does not make liaison in everyday talk but tries to make it in a formal setting produces some liaisons that the standard does not have. The studies cited here do not model it.

3.7 Liaison as a carrier of grammar

In spoken French the liaison consonant is a marker of number and gender that the spelling does not show. The regularity behind it is that the latent consonant of the masculine form is the final consonant of the feminine form:

The same reasoning applies to the plural: les /le/ before a consonant and /lez/ before a vowel is the only audible mark of number in les enfants. Each of the following has a number or gender mark in speech only in liaison: mon/ma (mon amie /mɔ̃.na.mi/, with the masculine form of the possessive before a feminine vowel-initial noun), un/une, des/les/ces/mes/tes/ses. This is the basis of the UD feature in the spoken treebank ParisStories (HasSpokenGender, HasSpokenNumber; §9). The effect on grammar is large: in conversation a speaker who does not make the liaison leaves the plural of des amis unspoken, and the listener relies on the verb or the context. It also explains why the liaison of the plural noun + adjective (des enfants intelligents) is not obligatory: the number has already been said.

4. The two h's

French has two classes of word with an initial written h. Neither is pronounced, except in a few interjections (hop! hé! ha ha!) and in some loans (hamburger with [h] in affected speech).

• Mute h (h muet): the word behaves as if it began with a vowel. It takes elision, enchaînement and liaison: l'homme, l'histoire, l'heure, les hommes /le.zɔm/, un homme /œ̃.nɔm/, cet hôtel, quel hôpital, ils habitent /i.la.bit/. These are the great majority of the h-words of Latin origin (homme, heure, histoire, hôtel, hôpital, honneur, humain, huile, hiver, horreur, habiter, hésiter).
• Aspirate h (h aspiré): the word behaves as if it began with a consonant. It blocks elision and liaison, and it forces a hiatus: le héros, la haine, le haricot, les héros /le.e.ʁo/, en haut, des hiboux, un hamster, quel hasard. The name is historical; the aspiration was lost in the standard centuries ago. Martinon, quoting the argument of the Académie that the letter keeps a role, notes that the h is a diacritic: "it is because one says il est hardi or des homards without liaison, and hence with hiatus, and le harnais or la hotte without elision, that the h plays its role" (martinon-1913). He also notes that the speech of the people tends to treat the aspirate h as mute ("if language were left to itself the aspirate h would soon become identical to the mute h").

Which words are aspirate. The aspirate words are a closed class of about two hundred lexical items, which the dictionaries mark with an asterisk, a dagger or the word "aspiré". The majority are of Germanic origin or expressive, and the Latin words with an aspirate h are those whose Latin h was reinforced by Germanic contact. Examples:

Not all derivatives behave like the primitive: héros is aspirate (le héros) but héroïne, héroïsme are mute (l'héroïne). Hiérarchie is aspirate in the dictionaries (la hiérarchie; GSD has la hiérarchie 7 times and no l'hiérarchie), though l'hiérarchie is heard; hiatus takes both (le hiatus, l'hiatus). Hyène is mute in the standard but aspirate in some uses; Martinon notes that "the word hyène is not in the same case as yacht" (the y of which acts as a consonant) and that La Fontaine treated hiéroglyphes as mute. The word huit is aspirate when it stands alone in a counting context but takes the liaison /t/ when followed by a noun (huit‿ans /ɥi.tɑ̃/ is accepted in speech).

Counts. In the UD French GSD treebank (UD 2.18, three files, counted 2026-11-06) the apostrophe form l' before an h-initial word is followed by mute-h words only: of the 7 384 tokens l' the h-initial forms that follow are homme (91), histoire (76), hôtel (25), hôpital (24), honneur (22), heure (13), humanité (12), huile (10), … Not one aspirate word. After the unelided articles le/la the h-initial forms that follow are hauteur (14), hausse (9), haut (8), hiérarchie (7), haye (7), haute (7), héros (6), havre (6), hasard (6), … The two sets are almost disjoint, which is the relevant fact for a tokeniser: in a written French text l' + h-word identifies a mute-h word.

5. Elision

Elision drops the final vowel of a small set of function words before a word that begins with a vowel or a mute h, and writes an apostrophe. It is the one process of this chapter that the standard spelling records; and it is obligatory in writing, whatever the speed of speech.

Where elision does not apply.

• before an aspirate h: le héros, la haine, le haricot, ce hibou;
• before oui, onze and the vowels of loan words that act as consonants: le oui, le onze, le yacht, le yoga, le ouistiti; but l'oiseau has the vowel /wa/ and elides (l'ouest, l'ouïe);
• before a vowel-initial word that is cited or titled: le ah! de surprise, and for numbers: le un (the numeral), de un à dix (Martinon: "de une heure à deux, without elision", since the numeral un is taken as a word and not as the article);
• after an imperative, the object clitic keeps its vowel in writing: fais-le entrer, laisse-le aller (not *fais-l'entrer), though the e is elided in speech;
• before the name of a letter or a word cited as such, usage varies: le a, le i (letters), le « oui » de la mariée.

Frequency. The elided function words are among the most frequent tokens of written French. In UD French GSD (UD 2.18, train + dev + test, 400 385 syntactic words, counted 2026-11-06) the elided function words are 16 478: l' 7 384, d' 5 028, s' 1 049, qu' 1 026, n' 823, c' 554, jusqu' 245, j' 237, m' 61, lorsqu' 50, puisqu' 15, t' 4, quoiqu' 1, presqu' 1 (the other forms that end in an apostrophe are 40 quotation marks typed as ' or '' and a few English and dialect forms such as swingin', y'). Quelqu'un (205) is one word in UD and not split. About one syntactic word in twenty-four is an elided form. The frequency of t' (4 tokens against 237 for j') shows how much the informal spelling is absent from the written corpus; the fuller picture of the spoken forms is in 37. Register and variation. Aujourd'hui (143 tokens) is a single word and is not an elision.

Elision in speech. The written elision matches the sound only for the vowels. In speech, many other schwas are also elided, with no apostrophe: je ne sais pas /ʒən.se.pa/ → /ʃɛ.pa/ (see §7). The elision of the final e of a polysyllable (une petite amie) is the mute e rule.

Frequency and elision. The rate at which a clitic loses its schwa varies by collocation. Liégeois (liegeois-2016) studied a corpus of natural interactions at home (six speakers) and found that the rate of schwa elision in "clitic + X" sequences (je + verb, ce + word, de + word, que + word, le + word) depends significantly on the usage frequency of the sequence: the more frequent the collocation, the more often the schwa is dropped. This confirms for monosyllabic clitics what earlier work had shown for initial schwas in polysyllables.

6. Contraction (amalgame)

Contraction is the fusion of the prepositions à and de with the definite article le or les, written as a single word:

A contraction is not made before a vowel (à l'ami, de l'eau). The article of a place name contracts like any other (au Havre, du Mans, aux Pays-Bas); the Le of a family name does not (les romans de Le Clézio); before a title, usage is divided (un article du Monde, or de Le Monde in some house styles). The contracted forms du and des are ambiguous with the partitive and the indefinite plural article: du pain ('some bread') against du pain ('of the bread'); des livres ('some books') against des livres ('of the books') (12. Determiners, pronouns and numerals).

UD treats the contracted forms as multiword tokens: the surface string au is a range line over two syntactic words à (ADP, case) and le (DET, det). In UD French GSD (all three files, counted 2026-11-06) there are 11 022 multiword-token range lines: du 4 390, des (contracted) 3 298, au 2 477, aux 808, auquel 20, duquel 15, auxquels 6, desquelles 5, auxquelles 3 (the 2026-10-22 counts in the earlier chapter are for training only). They are the single largest class of tokenisation mismatch between a whitespace tokeniser and the treebank (§9).

7. The mute e (schwa)

7.1 What it is

The letter e at the end of a word and in many unstressed syllables of the middle of a word (petit, samedi, vendredi, boulanger, appartement, je, le, que, de, ce, ne) stands for a vowel that is optional: it may be pronounced as /ə/ (for many speakers a rounded [œ] or [ø]), or may be absent. Passy 1906 and Martinon 1913 call it the e muet or e caduc (falling e); modern authors say schwa. Its status in the phonology is debated (is it a vowel with a zero alternant, or a floating position?); the corpus work below is neutral on this point.

The pronunciation of schwa depends on five kinds of conditions, which interact and are all gradient.

7.2 The law of three consonants

The oldest rule is the law of three consonants (loi des trois consonnes), associated with Maurice Grammont and his Traité pratique de prononciation française (1914; card only in our library): schwa is pronounced when its omission would produce a sequence of three consonants, and dropped when it would leave two.

The law is a good description of the tendency and an incomplete rule. Storme 2021 (storme-2021) tested it with acceptability judgments on stem-level words (derived with -rie, -ment, -erie) and word-level words (inflected forms such as il garde-ra), and shows that the nature of the consonants matters at both levels, not just the number of them: clusters differ in how well they tolerate a missing schwa, in the same direction in derived and inflected forms. The result fits a grammar in which the same phonotactic constraints (sonority, place) apply at all levels, which weakens the usual stem-level versus word-level divide. It also replicates earlier findings on clash avoidance (a schwa is more likely to be kept next to another schwa or a stressed syllable) and on dialectal variation. It does not overturn the law as a first approximation: three consonants in a row are much less acceptable than two.

7.3 Position and word type

• Final in a word before a consonant: une petite table /yn(ə).pə.tit(ə).tabl(ə)/. In running speech of the north and of Québec the final schwa is dropped in most positions; in the south of France, in verse and in careful reading it is pronounced (Passy and Martinon both point to this regional difference).
• Final in a word before a vowel: always elided (une amie /y.na.mi/); the e is silent and the consonant before it is enchaîné.
• Initial in a monosyllabic clitic (je, me, te, se, le, ne, de, ce, que): variable. After a vowel it is almost always dropped (tu le sais /ty.l(ə).se/), after a pause it is kept (Je sais, but */ʒsɛ/ is also heard, with a devoicing: [ʃsɛ]). The common spoken results are the subject of §7.5.
• Initial in a polysyllable (retour, demander, petit, venir, semaine): variable and strongly frequency-dependent (Liégeois, liegeois-2016).
• Medial: boulangerie /bu.lɑ̃.ʒə.ʁi/ against /bu.lɑ̃ʒ.ʁi/; appartement /a.paʁ.tə.mɑ̃/ against /a.paʁt.mɑ̃/.

7.4 Variation between and within speakers

The schwa is a textbook example of variation that is not categorical in the individual speaker. Bayles, Kaplan & Kaplan (bayles-2016) studied the optional deletion of schwa in a French corpus and found two things: (i) individual speakers do show variable omission, so the variation is not only the sum of speakers who always drop and speakers who never do; (ii) speakers also differ among themselves in how often they keep the schwa. They conclude that the variation must be placed in the grammar of a single speaker, and that the formal theories which decouple the means of producing variation from the account of output frequency are favoured. For description it means that rules such as the law of three consonants state a tendency by environment, and the actual rate depends on the speaker, the register, the speech rate and the word.

The two speech conditions that raise schwa pronunciation are slow rate and formality (reading aloud, the stage); the two that lower it are fast rate and the north (Paris, Québec, Belgium). The south keeps nearly all spelled e's, to the extent that the "southern accent" is described by outsiders as the pronunciation of every e.

7.5 The usual spoken results

Frequent combinations produce standard reduced forms, regularly written in informal text and in the transcription of speech:

The spelling of these forms in writing (chais pas, chuis, t'es, y a, j'sais) is a device of informal registers and of the transcription of speech; the manual's register chapter (37. Register and variation) counts them.

7.6 The schwa in verse

The conventions of French classical verse are a codification of the mute e of the seventeenth century. The syllable count of a line counts every e before a consonant as a syllable (une petite fille has seven syllables inside a line of verse and three or four in conversation), does not count the e before a vowel (elision: la jeune amie → la jeun'amie), and does not count a final e at the end of a line before the pause (the "feminine" ending, which gives the rime féminine). Within the line, the e after a vowel (vie, joue) before a consonant is a problem the poets solved by forbidding it; and hiatus (a vowel followed by a vowel) is forbidden since Malherbe, with a few exceptions. Kastner (kastner-1903) notes that "the methods followed by French prosody for the counting of the feminine e in verse are in contradiction with the present pronunciation even of cultured society". Verse thus keeps, as a convention, the full pronunciation that the northern speech had lost. For a model that wants to scan verse, the spelling alone gives the syllable count except for four things: the e before a vowel word (silent), the e after a vowel (silent), the e at the line end (silent), and the diaeresis/synaeresis of vowel-vowel sequences within a word (lion, nation, pi-é-tà, which Kastner treats at length): the last two require a word list.

8. Assimilation

Assimilation is the influence of a sound on the one next to it. In French the main domain is voicing; the second is nasalisation; place assimilation is minor.

8.1 Voicing assimilation

A consonant takes the voicing of the following consonant (regressive assimilation), and sometimes of the preceding one (progressive). Nyrop (nyrop-1902, §§153–156) gives the classical examples and notes that it is "impossible to say by rules when assimilation occurs and when not". His examples of voiced → voiceless (the more frequent direction):

The opposite direction (voiceless → voiced):

Voicing assimilation is gradient: it can be partial and affect only the start or end of the consonant (Nyrop). It is the French behaviour that makes médecin /mɛt.sɛ̃/ and absurde /ap.syʁd/ the regular pronunciations, and it can be written into the spelling of the informal style: chuis, chais pas. The same phenomenon explains the pairs obtenir /ɔp.tə.niʁ/ with /p/, subsister /syp.sis.te/ against /syb.zis.te/ (Nyrop notes the verb "has at present two pronunciations and it is very difficult to decide which is the better"), anecdote /a.nɛɡ.dɔt/, rugby /ʁyɡ.bi/.

Progressive assimilation is found when a voiceless consonant devoices the following liquid or glide: clef /kle/ with a voiceless [l̥], pied /pje/, prêt [p̥ʁɛ], quoi [kʷwa], tuer [tɥe] with devoiced [ɥ̥], enthousiasme /ɑ̃.tu.zjasm/ → [asm̥] (the voiceless /s/ devoices the final /m/): the voiceless articulation propagates (Nyrop §§153–154). It is more phonetic detail than phonemic change.

A corpus view. Vasilescu et al. (vasilescu-2020) measured voicing alternations in about 981 hours of media speech in five Romance languages, aligning canonical against non-canonical pronunciation variants with an automatic speech recogniser. They found that the stops /p t k/ are more apt to be realised voiced (lenition) in word-internal position and that /b d ɡ/ are more apt to be realised voiceless (fortition) at word-initial position — a result that holds across French, Spanish, Italian, Portuguese and Romanian, and which the authors set beside the historical lenition of intervocalic stops (Latin focus > feu), one of the classic criteria for dividing the Romance languages.

8.2 Nasalisation, place and degemination

• Nasal assimilation is mostly historical: the oral vowel + nasal consonant sequences of Latin became nasal vowels (Latin cantare > chanter /ʃɑ̃te/), in Nyrop's account the nasalisation of the vowel in anticipation of the nasal consonant, then the loss of the consonant (nyrop-1902). In the living language it is present at word junctions in the denasalisation of bon ami /bɔ.na.mi/ (§3.1), where the nasal vowel of bon gives way to oral [ɔ] + /n/ before a vowel and stays nasal before a consonant (un bon pain [œ̃.bɔ̃.pɛ̃]).
• Place assimilation is rare in French: a nasal vowel has no consonant to assimilate (un pain [œ̃.pɛ̃], en bas [ɑ̃.bɑ], not *[ɑ̃m.bɑ]), and a final /n/ or /t/ keeps its place before a labial or velar (une bonne place [bɔn.plas]), unlike English ten boys [tem bɔɪz]. The spelling itself carries a place assimilation: m before b and p (lampe, timbre, emmener) with the nasal vowel already marked.
• Degemination: a double consonant at a boundary or inside a word is often pronounced as one, except when the double letter is the sign of a distinction (il courrait [ku.ʁːɛ] against il courait [ku.ʁɛ], nous mourrons against nous mourons): il l'a [il.la] keeps two /l/ only in careful speech, là-dedans [lad.dɑ̃] ~ [la.dɑ̃]. The double rr of the conditional and future is a real contrast in careful speech.
• Linking and vowel hiatus: a hiatus between two vowels across a word is avoided by liaison, enchaînement or a glottal stop; inside a word, vowel sequences are separated by a glide (lier /lje/) or by a syllable break (créer /kʁe.e/, Noël /nɔ.ɛl/). The diaeresis in the spelling marks the second case.

9. How Mova handles connected speech

What is done. Mova works from the written form. It has no phonological component for French: no grapheme-to-phoneme rules, no liaison model, no syllabification, no schwa model, no assimilation. French goes through the English parser (en) with tags mapped by en xpos-fill (README). Of the processes of this chapter, only elision and contraction exist as written forms, and Mova's tokenizer is the English one.

What the English tokenizer does with French elision (en/src/tok.rs). The tokenizer is a table-driven automaton: a letter followed by an apostrophe (class Ap, which covers ' and ’) followed by a letter stays inside the word. So l'arbre, qu'il, jusqu'alors, aujourd'hui come out as one token each, whereas UD French cuts the elided word off (l' + arbre, jusqu' + alors) and keeps aujourd'hui whole (143 tokens in GSD, counted 2026-11-06). The splitting of clitics is done by a list of English endings (n't 's 're 've 'll 'd 'm) and there is none for French. Hyphenated words are kept or split according to a dictionary learned from the training treebank; for GSD it learns 1 994 hyphenated words kept whole against 46 split, so the default is to keep: Va-t-il stays as one token, where UD has Va + -t-il. French multiword tokens (au, du, des, aux, auquel) are never split: the English tokenizer has no notion of them, and conllu::read trains on syntactic words, so a raw French text yields au where the model expects à + le. The French non-breaking and narrow spaces are Unicode whitespace for the automaton, so a number written 25 785 with a non-breaking space is cut in two (not tested on raw text; by the code); guillemets are single punctuation tokens.

Measurement. en tok-eval with the dictionary learned from the GSD training file, tested on the GSD development and test files, on the gold syntactic tokens (the tool compares the system tokens with the treebank's words), 2026-11-06 (binary built 2026-10-22; two runs gave the same figures):

The misses printed by the tool are exactly the three forms above: L' | oasis, l' | obscurité, jusqu' | alors, de | les for des, à | les for aux, C' | est. A French tokenizer with an elision rule, the clitic hyphens and the contraction table is the first piece of the French stack to write (07. Punctuation, typography, capitals, abbreviations and numbers); until then Mova's French parsing numbers are measured on gold tokens, which is what the LAS figures of the README do.

Spoken-language data that Mova does not read. The open spoken treebank UD French ParisStories (UD 2.18, all three files, 2 776 sentences, 42 789 syntactic words, counted 2026-11-06) carries phonological information in its MISC column:

Each of the 526 LiaisonAfter tokens is followed directly by a LiaisonPossibleBefore token; the other 240 of the 766 vowel-initial tokens with a possible liaison are not preceded by a marked one. The README of the treebank does not define the two features, so we read LiaisonAfter as the annotation of a liaison that was made and LiaisonPossibleBefore as the context in which one is possible (an interpretation of the data, not a statement of the annotators). Mova's en/src/conllu.rs reads the MISC column only for SpaceAfter; the liaison and spoken-gender features are ignored. This treebank is the only open source known to us that links written tokens to liaison behaviour, and the natural first use is a table of liaison contexts and rates by word pair.

Spelling to sound. There is no French pronunciation lexicon in Mova and no model of the latent consonant: the information "this word has a latent t" exists nowhere in the data except in the spoken treebank's features. A dictionary with the latent consonant of each lexeme (the fifteen or so classes of §3) plus the list of aspirate h words (§4) would be enough to produce liaison candidates; the register-dependent decision (§3.4) needs corpus rates by word pair, as in Bosch & Greefhorst. These are possible additions, not existing parts.

What is not handled. No phonemes, no liaison, no enchaînement, no schwa deletion or syllable counting (needed for verse and speech synthesis), no assimilation, no handling of informal spelling of spoken reductions (chais pas, t'es: the tagger sees them as unseen words; the tagger's accuracy on words it has not seen was 84.8% on GSD development, 2026-10-22).

Sources for this chapter

• Bosch, Antal van den & Alexander Greefhorst. 2016. Predicting Liaison: an Example-Based Approach. TAL 57(1): 13–32. CC BY 4.0 (ACL Anthology). bosch-2016; aclanthology.org/2016.tal-1.1
• Storme, Benjamin. 2021. Not only size matters: limits to the Law of Three Consonants in French phonology. Glossa 6(1): 125. CC BY 4.0. storme-2021; doi.org/10.16995/glossa.5882
• Liégeois, Loïc. 2016. Effet de la fréquence d'usage sur l'élision du schwa des clitiques : étude d'un corpus d'interactions naturelles. JEP-TALN-RECITAL 2016. CC BY 4.0 (ACL Anthology). liegeois-2016; aclanthology.org/2016.jeptalnrecital-jep.28
• Bayles, Andrew, Aaron Kaplan & Abby Kaplan. 2016. Inter- and intra-speaker variation in French schwa. Glossa 1(1): 19. CC BY 4.0. bayles-2016
• Vasilescu, Ioana, Yaru Wu, Adèle Jatteau, Martine Adda-Decker & Lori Lamel. 2020. Alternances de voisement et processus de lénition et de fortition : une étude automatisée de grands corpus en cinq langues romanes. TAL 61(1): 11–36. CC BY 4.0 (ACL Anthology). vasilescu-2020
• Martinon, Philippe. 1913. Comment on prononce le français. Paris: Larousse. Public domain. martinon-1913
• Nyrop, Kristoffer. 1902. Manuel phonétique du français parlé, 2nd ed., trans. E. Philipot. Public domain. nyrop-1902
• Passy, Paul. 1906. Les sons du français, 6th ed. Public domain. passy-1906
• Kastner, L. E. 1903. A History of French Versification. Oxford: Clarendon. Public domain. kastner-1903
• Grammont, Maurice. 1914. Traité pratique de prononciation française. (Known to us by its card and through Storme 2021 and Liégeois 2016; not read.) grammont-1914
• UD French GSD, UD French ParisStories, UD 2.18 (CC BY-SA 4.0); counts of 2026-11-06 from UD_French-GSD and UD_French-ParisStories/.
• Mova: en/src/tok.rs, en/src/conllu.rs, en/src/main.rs (tok-eval, xpos-fill); measurements of 2026-11-06.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/fr/04-connected-speech/