How accurate is French speech to text with local Whisper?
French is one of Whisper's strongest languages, grouped with English, Spanish, and German at the top. On clean audio the largest model lands near English-level accuracy. Everyday dictation in a normal room runs higher than a clean benchmark, but French is firmly production-grade on a good local model.
Accuracy scales with model size, not with local-versus-cloud - it is the same open-source Whisper. French improves steadily from medium to large to large-v3; the compact free-tier models degrade more on French's accents and elisions than on English. SnailText's larger French-quality models are on Pro.
One thing worth knowing: fine-tuning a Whisper model on French is not a guaranteed win - French-fine-tuned variants have been shown to do worse than the base model on Québec French. That is why SnailText does not claim a fine-tuned French model. Instead it keeps the strong base model and adds a targeted prompt and cleanup pass, which addresses the documented weakness without the fine-tuning downside.
The French homophone problem, and what we do about it
French's hardest speech-to-text problem is not accents - Whisper outputs é è ê ç and elisions like qu'est-ce que c'est correctly. The hard part is homophones: words that sound identical but are spelled differently and mean different things. a (has) vs à (to), ou (or) vs où (where), c'est vs ses vs ces. When Whisper lacks the context to choose, it picks wrong, and the transcript reads as a spelling mistake even though the pronunciation was heard perfectly.
This is well documented in the research on Whisper and French, and it is exactly what SnailText targets. The French cleanup pass looks at context and fixes the homophone that is clearly wrong - il a fini versus il va à, for instance - so the output reads as correct written French, not a phonetic transcript.
Accents and elisions come out right from the model itself, so the app's job here is disambiguation, not inserting diacritics. Regional French is the honest limit: Metropolitan (France) French transcribes noticeably better than Québec French across every tool, and that gap is a property of the audio, not the app.
Why local, and why French-tuned
Most dictation tools ship stock Whisper and leave the homophone problem untouched. SnailText is the combination that fixes it privately: the same best-in-class local Whisper model everyone runs, plus a French-tuned prompt and a French homophone-aware cleanup pass, running entirely on your own hardware.
That matters when your French dictation is sensitive - legal drafts, medical notes, client work, internal documents. Because the audio is processed in RAM and never uploaded, the privacy guarantee is the architecture, not a policy you trust. It also works with no internet at all.
Want the interface in French too? SnailText ships a fully French UI. The dictation engine is the same; these pages are about dictating in French whatever your interface language.