SnailText
EN

French speech to text

French speech to text, running on your own machine

·

Dictate in French into any app on Mac or Windows. SnailText runs Whisper locally - one of its strongest languages - with a French-tuned prompt and a homophone-aware cleanup pass for the a/à, ou/où, c'est/ses/ces traps. Nothing is uploaded.

Download for Macand start dictating in any app

The short version

French is one of Whisper's best-supported languages: on clean audio the large model lands in the top tier, close to English. Its accents and elisions come out right - the genuinely hard French problem is homophones, words that sound identical but are spelled differently (a vs à, ou vs où, c'est vs ses vs ces). Stock Whisper picks the wrong one when it lacks context. SnailText runs the model locally on Mac and Windows, adds a French-specific initial_prompt, and runs a French cleanup pass that targets exactly those homophones. Press a hotkey, speak French, and the text is pasted at your cursor. Audio is processed in RAM and never leaves your device.

French dictation: local Whisper vs cloud STT

SnailText (local)Typical cloud French STT
Where audio goesStays on your device, in RAMUploaded to a server for every phrase
Works offlineYes, after the model downloads onceNo, needs a connection every time
Accents é è ê ç and elisionsPreserved by the modelUsually preserved, varies by vendor
Homophones a/à, ou/où, ces/sesCleanup pass disambiguates by contextLeft as the model guessed
French-specific tuningFrench prompt + homophone-aware cleanupMost run stock Whisper with no per-language layer
Account / costNo account, free tier, unlimitedAccount + per-minute or per-seat billing

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 (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.

Talk instead of typing

Download for Mac

and start dictating in any app

Frequently asked questions

How accurate is French speech to text?

+

On clean audio, local Whisper reaches top-tier accuracy on French with the large model - one of its strongest languages, close to English (Whisper's English word error rate on the same benchmark is around 2-3%, and French is in the same top tier). Accents and elisions come out right. The genuinely hard part is homophones (a/à, ou/où, c'est/ses/ces), which SnailText's French cleanup pass disambiguates by context.

Can I use this to transcribe French audio, or is it only live voice to text?

+

The core of SnailText is live French voice to text: press a hotkey, speak, and the text is pasted at your cursor in any app. It is built for dictating as you go rather than batch-transcribing existing recordings, though the same local Whisper model powers both. Turning French audio files into text is a related but separate workflow.

Does French dictation work offline?

+

Yes. SnailText runs the Whisper model on your own Mac or Windows machine, so French dictation works with no internet connection after the model downloads once. Audio is processed in RAM and never uploaded.

Does it fix French homophones like a/à and c'est/ses/ces?

+

That is the specific thing SnailText targets for French. Whisper hears the pronunciation correctly but picks the wrong spelling when it lacks context. SnailText runs a French homophone-aware cleanup pass that looks at context and corrects the clearly-wrong choice - a vs à, ou vs où, c'est vs ses vs ces - so the output reads as correct written French.

Does it handle accents and elisions?

+

Yes, and those come directly from the model - Whisper outputs é è ê ë à â ç and elisions like qu'est-ce que c'est correctly. The app's French-specific work is disambiguating homophones, not inserting diacritics, because the diacritics are already right.

Does it work with Québec French?

+

It works, but Metropolitan (France) French transcribes noticeably better than Québec French - a gap that shows up across every speech tool, and one that French fine-tuning does not reliably close (fine-tuned models have done worse on Québec French than the base model). We flag this rather than imply the accuracy is uniform.

French speech to text, on your own machine.

Free to start on Mac and Windows. Press Option+Space (Mac) / Ctrl+Space (Windows), speak French, and the text lands at your cursor - homophones sorted out. No account, nothing uploaded, works offline.

Download for Macand start dictating in any app