What happened
Mistral AI has released Voxtral 3, speech models for transcription and text-to-speech. Its announcement says they are available through Mistral's EU-hosted API or for self-deployment, with US and global endpoints alongside.
Voxtral Mini TTS 3 speaks more than 35 languages, with preset voices in 18, including French, German and Spanish. A Pro Voice option builds a custom voice from 30 minutes of recorded audio. Speech is listed at US$13 per million characters.
Why it matters
For European editors and producers, the endpoint decides where an actor's or narrator's recordings are sent. Voxtral 3 can keep that work on Mistral's European servers or in-house, which helps when a contract limits where material may travel.
At about six characters per word, a 220-word voice-over runs to 1,300 characters, so 12 language versions cost roughly US$0.20 at the listed rate. As of 28 October, these terms are on Mistral's announcement page.
Still open
The announcement does not say whether Voxtral 3 weights are open or which licence covers self-deployment, and it states no rule on cloning a voice without the speaker's consent. Mistral's text-to-speech documentation, which names the model only as Voxtral TTS, describes cloning from 2 to 3 seconds of audio, against 30 minutes on the Voxtral 3 page. The page gives no EU-endpoint price and no publication date; its charts are dated 20 October.