Voxtral TTS: what a French, open speech model changes
12 October 2026 · 7 min read · Figures checked on 12 October 2026
On 23 March 2026, Mistral AI released Voxtral TTS, its first text-to-speech model. Two things stand out: the lab is French, and the model weights can be downloaded. If you produce voice-overs, podcasts or training modules, that changes which tool you pick, what a minute costs and how you read the usage rights. Here is what matters, keeping what Mistral claims separate from what is verified.
What Mistral states, in black and white
The facts below come from Mistral's announcement and pricing page, checked on 12 October 2026. Voxtral TTS has about 4 billion parameters. It speaks nine languages: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic. The hosted model is called voxtral-mini-tts-latest and costs $0.016 per 1,000 characters, or $16 per million. It accepts a three-second reference clip to imitate a voice, and Mistral quotes a model latency of 70 ms for a typical ten-second sample.
On quality, Mistral writes that Voxtral TTS "achieves superior naturalness compared to ElevenLabs Flash v2.5" and "quality parity with ElevenLabs v3", with a 68.4% preference rate in human evaluation. Those are the vendor's own tests, not an independent measurement: as we write, no third-party benchmark confirms or contradicts them. The practical answer is to listen to the same text on several voices, which the voices page lets you do, then run the free trial on your own script.
Open weights and API are not the same thing
An "open weights" model can be downloaded and run on your own hardware. Voxtral TTS weights sit on Hugging Face under a CC BY-NC 4.0 licence. The "NC" means non-commercial: you may test, learn and produce for yourself, but not sell a service or ship the model inside a paid product without an agreement with Mistral. You also need a decent GPU and someone who knows how to deploy a model, which rules out most creators.
The API is the commercial route: you pay per character and you may publish, monetise and deliver to clients. That is the path services like Aievalu use. If you are wondering what a generated audio file allows you to do, the short answer is in our FAQ on usage rights; the long answer depends on the chain: non-commercial weights on one side, API and licensed services on the other.
Why a French lab changes what you hear
Most speech engines were designed for English first and extended afterwards. French often arrives with missing liaisons, numbers read the English way and intonation rising in the wrong place. A model trained by a team that speaks French every day is no guarantee, but it is a good signal: French prosody, breathing commas and figures such as "1 500", "2026" or "9 €" are treated as a primary case, not a translation. Our own tests on 1,500-character passages point the same way, with one limit we detail in the guide to text to speech with natural voices: punctuation remains your only directing tool. For English, the American and British presets carry their own accent and rhythm, which matters for a UK audience.
European hosting and texts that are not kept
A French model called from a French API also means a processing chain that can stay inside the European Union. On Aievalu, the texts sent to the engine are not stored: only a technical fingerprint is kept, and the audio stays 90 days in your history before deletion. For a company whose legal team reviews scripts, that weighs as much as the price.
What a minute actually costs
At $16 per million characters, a minute of speech (about 1,000 characters) costs 1.6 cents through the API, before any development work. Consumer platforms charge the same minute between $0.17 and $0.24 on their entry plans, with a monthly subscription attached. Aievalu sits in between: a 100,000-character pack for €9, valid twelve months, about €0.09 per minute, no subscription. The tool-by-tool detail, with a calculator, is on the AI voice price comparison, and the model-versus-model view on Voxtral vs ElevenLabs.
What Aievalu does with Voxtral, and what it does not
Aievalu uses Voxtral TTS preset voices: Marie in French (three tones: neutral, excited, happy), Paul in American English, Oliver and Jane in British English. You paste a text, pick a tone, listen, download as MP3 or WAV. Every MP3 carries a technical tag marking it as synthetic speech, which European rules now require.
The limits are just as clear. No voice cloning: we do not expose the reference-audio feature, by choice, until consent can be handled simply. Only two languages, French and English, while the model speaks nine: we open a language only after listening to it seriously. And one generation is capped at 1,500 characters, roughly ninety seconds, so long texts must be split, as explained in the podcast voice-over guide. If you need cloning or thirty languages, our round-up of the best AI voice generators will tell you where to go.
Who this model really changes things for
For a YouTube or TikTok creator publishing in French, "correct French prosody plus per-character pricing" removes the main argument for a $22 subscription. For a small business voicing training modules, European hosting and non-retention of texts make internal approval easier. For a developer, an API at $16 per million is among the cheapest on the market, provided you are willing to write code. And for a researcher, open weights finally allow a synthesis model of this level to be studied without a black box.
The best way to form an opinion takes two minutes: paste a paragraph of your next script into the free trial on the home page, listen, then compare with your current tool on the same text. Full prices are on the pricing page.