French Speech to Text
Liaisons, accents, and Quebec French

French transcription is hard for general speech recognition models because so much of the language depends on context: the same sound can be three different words (vert / verre / vers), liaisons connect words across boundaries (les-z-amis), and dropped e and t markers carry grammatical meaning. Whisper large-v3, which Mictoo runs, was trained on a substantial French corpus and handles these patterns well.

Useful for French journalists transcribing interviews, university professors recording lectures, content creators shipping French-language podcasts, businesses operating in Quebec or France, and anyone working with French audio content that needs to become text.

The upload form is pre-set to French for the cleanest first attempt. Auto-detect works for files longer than 30 seconds, but explicit French selection is more reliable for short clips or files that open with music or silence.

What makes French speech recognition difficult

French is harder than English or Spanish for generic speech models because of three patterns that English-trained architectures historically struggled with: liaisons, élisions, and the e-muet (silent e). All three move phonological boundaries away from word boundaries, which is the opposite of what a naive word-segmentation model expects.

Liaisons

A liaison is when the normally-silent final consonant of one word becomes audible to link with the vowel that starts the next word. "Les amis" pronounced /le.za.mi/, with the s of "les" appearing as a z linking into "amis". "Vous avez" pronounced /vu.za.ve/. There are obligatory liaisons (after determiners), forbidden liaisons (across major syntactic breaks), and optional liaisons (depending on register). Whisper large-v3 was trained on enough French speech to learn which sounds belong to which words even when the phonological break is in the middle of an apparent syllable.

Élisions

Élision is when a vowel drops to avoid hiatus. "Je aime" becomes "j-aime", "le ami" becomes "l-ami", "que est-ce que" becomes "qu-est-ce que". Spelled with an apostrophe, the audio just sounds like one fused syllable. Whisper correctly renders these as the conventional apostrophe forms rather than as concatenated single words.

The e-muet (silent e)

The letter e at the end of words is mostly silent in standard French ("table" pronounced /tabl/). But in poetry, songs, and some southern French varieties (Marseille, Toulouse), the e is pronounced ("table" pronounced /tablə/). The transcription has to render "table" both times even though the audio is different. Whisper does this correctly because the training data covered both varieties.

Regional varieties of French

Hexagonal French (the variety spoken in France) is what most ASR training data leans toward. Quebec French has different vowel inventories (more diphthongs, the distinctive ɛɪ̯ in "fait"), different vocabulary, and the tu-vous register shifted toward more "tu" use. Belgian French uses "septante" and "nonante" for 70 and 90 (vs Hexagonal soixante-dix and quatre-vingt-dix). Swiss French mostly aligns with Hexagonal. African French varieties (Senegal, Côte d-Ivoire, DRC) bring additional vocabulary and prosody.

Whisper large-v3 handles all of these, with occasional quirks: Quebec-specific vocabulary may get rendered in Hexagonal equivalents, very thick regional accents (rural southern France, deep Quebec joual) may have lower accuracy. For these cases, the inline editor lets you fix terms before exporting.

Numbers, dates, and times in French

Numbers are an area where French varies regionally and where transcription accuracy matters for journalists and researchers. Hexagonal French uses "soixante-dix" for 70, "quatre-vingts" for 80, "quatre-vingt-dix" for 90. Belgian and Swiss French use "septante" and "nonante" (and "huitante" in some Swiss regions for 80). Whisper handles both, but the transcript follows whichever variety the speaker used. If you need consistency across sources, fix in the editor before exporting.