French is hard for speech recognition in an unusual way: much of its spelling is not pronounced. Plural endings, verb persons and gender agreement are often silent, so a model can hear every sound correctly and still write the wrong word. Liaison and elision blur word boundaries, and a large set of homophones must be resolved from grammar alone.
Key takeaways
- Many French errors are spelling errors with correct sound: il chante and ils chantent are pronounced identically.
- Liaison and elision move consonants and vowels across word boundaries, so the audio does not segment where the text does.
- French numbers from 70 to 99 are built arithmetically (quatre-vingt-dix-sept = 4 × 20 + 17), and Belgium and Switzerland use different words.
- Formatting uses a decimal comma and a space as thousands separator: 1 250,50 €.
- Reson8's published FLEURS French WER is 4.71%, the lowest in its published comparison for French.
What makes French hard#
Silent letters and invisible grammar#
French orthography preserves distinctions that speech lost centuries ago. In the present tense, je parle, tu parles, il parle and ils parlent all sound the same. The infinitive parler, the participle parlé, the imperative parlez and, for many speakers, the imperfect parlait collapse onto one or two vowel sounds. Plural nouns and adjectives usually add a silent -s: la grande maison and les grandes maisons differ audibly only in the article.
An acoustic model cannot recover these endings from the signal. The language model has to infer them from agreement with a subject or article that may be several words away. Errors here are frequent and look careless to readers: les résultats sont arrivés transcribed as les résultats sont arrivé. WER counts it as one substitution; for a lawyer or a journalist it is an error that needs correcting.
Homophones#
French has an unusually large set of short homophones that differ in meaning and grammar.
| Sound | Possible spellings | Meanings |
|---|---|---|
| /sɑ̃/ | cent, sans, sang, s'en | hundred, without, blood, "of it" (reflexive) |
| /vɛʁ/ | vert, verre, ver, vers | green, glass, worm, towards |
| /sɛ/ | c'est, s'est, ses, ces, sait | it is, has (reflexive), his, these, knows |
| /a/ | a, à | has, to |
| /u/ | ou, où | or, where |
A typical output error is il s'est trompé written as il c'est trompé. Short-context models, or streaming models that must commit early, make more of these than models that see the whole sentence.
Liaison and elision#
In les amis, the silent s of les is pronounced as /z/ and attached to the next word, so the audio sounds like lé zami. Liaison is obligatory in some contexts, optional in others, and forbidden in a few (et never links). Its presence helps: it tells the model a plural or a vowel-initial word follows. Its variability hurts: speakers differ, and formal speech uses more of it than casual speech.
Elision drops a vowel before another vowel and writes an apostrophe: le hôpital becomes l'hôpital, je ai becomes j'ai. Scoring must decide whether l'hôpital is one token or two. Different normalisers answer differently, which moves French WER in ways unrelated to recognition quality. See word error rate for why the normaliser matters.
Casual reductions#
Spoken French routinely drops the negative particle ne (je sais pas for je ne sais pas), reduces il y a to something like y a, and shortens tu es to t'es. A reference that restores the standard form and a hypothesis that writes what was said will disagree, even though both are defensible. Decide upfront whether your transcripts should be verbatim or normalised.
Regional varieties#
French is spoken as a first language in France, Belgium, Switzerland, Luxembourg, Monaco and Canada, and as an official or widely used language across West and Central Africa. The main varieties that matter for ASR:
- Metropolitan French with regional accents. Southern French keeps final schwas (une petite with audible final vowels) and a nasal consonant after nasal vowels, which shifts acoustics away from Parisian norms.
- Belgian and Swiss French: close to metropolitan French in writing, with distinct number words and some vocabulary (déjeuner for breakfast in Belgium, where France says petit-déjeuner).
- Québécois French: affrication of /t/ and /d/ before /i/ and /y/ (tu dis sounding like tsu dzis), diphthongised long vowels, and many English loans. Models trained mostly on European French show higher error rates here.
- African French: a broad set of accents shaped by local languages, underrepresented in most public data relative to the number of speakers.
For how this affects model accuracy generally, see accents and regional speech.
Numbers, dates and formatting#
French numbers are a classic inverse text normalisation problem. From 70 to 99, France counts in twenties: soixante-dix (60 + 10) for 70, quatre-vingts (4 × 20) for 80, quatre-vingt-dix-neuf (4 × 20 + 19) for 99. A model that splits these at the wrong hyphen produces nonsense such as 4 20 19 for quatre-vingt-dix-neuf.
| Number | France | Belgium | Switzerland |
|---|---|---|---|
| 70 | soixante-dix | septante | septante |
| 80 | quatre-vingts | quatre-vingts | huitante (in some cantons) |
| 90 | quatre-vingt-dix | nonante | nonante |
Written conventions:
- Amounts: 1 250,50 €, with a space (strictly a narrow no-break space) as thousands separator, a decimal comma and the currency symbol after the number, separated by a space.
- Dates: 23/09/2026 or le 23 septembre 2026. The first of the month is an ordinal: le 1er octobre.
- Times: 14 h 30 in formal French writing, 14:30 in software.
- Punctuation: a space before ? ! : ; in France, not in Switzerland or Québec. It sounds minor until a customer rejects captions for it.
Data availability and public test sets#
French is well resourced. Public sets include:
- FLEURS (French split): read Wikipedia sentences, parallel with 101 other languages.
- Common Voice: one of the largest non-English subsets, crowd-sourced read speech.
- Multilingual LibriSpeech: around a thousand hours of French audiobooks.
- VoxPopuli: European Parliament speech, including French spoken by non-native MEPs.
- ESTER and ETAPE: French evaluation campaigns on broadcast news and TV debates, distributed through ELRA rather than freely.
Most of this is read or prepared speech from Europe. Spontaneous telephone French and non-European varieties are thin, which is where public numbers are least predictive. See ASR benchmarks.
How Reson8 approaches this#
French is one of Resonant-1's nine supported languages. On FLEURS French it scores 4.71% WER, the lowest in the comparison published in Reson8's benchmarks post. FLEURS is read speech from European French speakers, so test on your own audio before assuming the same figure for Québécois or telephone French.
Common questions#
Why does French ASR get plural and verb endings wrong?
Because they are usually silent. The model has to infer them from grammatical agreement, not from sound, so errors appear even when every phoneme was recognised.
Does one model work for France, Belgium, Switzerland and Canada?
Usually, with some loss on Québécois and African French accents. Number words and punctuation conventions differ, so configure ITN per locale.
How should l'hôpital be tokenised for WER?
Either way works if you are consistent. Split or keep apostrophe forms identically in reference and hypothesis, and state the choice when reporting results.
References#
- Conneau, A. et al. (2022). FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. arXiv:2205.12446.
- Ardila, R. et al. (2020). Common Voice: A Massively-Multilingual Speech Corpus. LREC 2020.
- Pratap, V. et al. (2020). MLS: A Large-Scale Multilingual Dataset for Speech Research. Interspeech 2020.
- Wang, C. et al. (2021). VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. ACL 2021.
- Galliano, S. et al. (2005). The ESTER Phase II Evaluation Campaign for the Rich Transcription of French Broadcast News. Interspeech 2005.
- Radford, A. et al. (2022). Robust Speech Recognition via Large-Scale Weak Supervision (Whisper). arXiv:2212.04356.