Voxtral 3: a voice that can stay in Europe

On 20 October 2026, Mistral introduced Voxtral 3: a family of models that turns speech into text and text into speech, plus an SDK for voice agents. It is meant to run on Mistral's API in Europe or on your own infrastructure. Voice cloning is part of the package.

Voices have long been part of our productions, and today they are made in the cloud. In Rapido, a spec ad for Electrolux, the German version is spoken by a real voice actor; we dubbed the Swedish version with ElevenLabs. In 21 ZUAN, a spec ad for Seagull 1963, the grandmother's voice comes from ElevenLabs in English and Chinese. And in When Everything Crumbles for the University of Bern, the director's voice was cloned with ElevenLabs from his own recordings, with his consent.

With a fictional character, the question of where a voice is computed is mostly technical. With a real person, it isn't. That's why we're taking a close look at Voxtral 3. We have not used or tested it yet.

What is actually new

Mistral ships four models under one name:

• Transcription: Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 (4 billion parameters, according to Mistral) for files, with speaker diarization and context for technical terms. 50 languages, including the 24 official languages of the EU.
• Realtime: Voxtral Mini Transcribe Realtime 3, with under 500 milliseconds of latency according to Mistral, and detection of when someone has finished speaking.
• Speech generation: Voxtral Mini TTS 3 with 35 languages and more than 50 dialects, steerable through instructions on emotion, pacing and emphasis.
• Voice agent SDK: built on LiveKit, it connects listening, reasoning and speaking in one pipeline and hooks into phone lines via SIP or into browsers and apps via WebRTC. Mistral quotes under one second per conversational turn.

What is new is mainly the combination: one provider for the whole path from voice in to voice out, with EU hosting as a standard option and self-deployment as another. According to the announcement, the full stack on-premises or in a private cloud is available by talking to Mistral's team. Worth noting: besides the European endpoint, Mistral also lists US and global endpoints. Europe is a choice, not a default.

The benchmarks in the announcement come from Mistral itself, some of them judged by a language model. We leave them where they are.

What it means for a production

Mistral describes two ways to clone. Zero-shot clones from a short recording. For the professional option ("Pro Voice"), the docs recommend 30 minutes to three hours of clean audio with a single speaker: quiet room, same microphone, a range of emotions. According to the docs, a cloned voice is not tied to the language of its recording. That is exactly our situation with versions in other languages.

Prices according to Mistral: speech generation 13 dollars per million characters, transcription 0.0035 to 0.008 dollars per minute, the voice agent 0.04 dollars per minute with Mistral Small 4 as the reasoning model. For a commercial, that's not a cost factor. For an agent taking calls all day, it starts to be one.

What it means for a brand

This is where it gets interesting. According to Mistral, organisations can control where audio is processed, how data is handled and whether conversations are retained. Saved voices have a retention period according to the docs, 30 days by default, adjustable up to 365 days.

For comparison, here is what ElevenLabs documents: customer data is stored in the US by default. Storage in the EU is a feature for Enterprise customers, and according to the docs, processing may still happen outside the region unless you use Zero Retention Mode through the API.

On consent, it's the other way round. According to its docs, ElevenLabs only allows professional cloning of your own voice and requires verification. Mistral's cloning docs describe no such step. Its usage policy does, however, prohibit using another person's voice without their prior consent.

Meaning: consent is on you, not on the tool. Honestly the most important point in this post.

So when is a voice that stays in Europe, or runs on your own infrastructure, worth it? When it belongs to a real person, say the CEO, a voice actor or a director. When the material is confidential before launch. Or when a voice agent talks to customers and hears personal data along the way. For a fictional character in a spec ad, it matters less. The legal side has to be assessed case by case by a specialist; what we can share is the questions we ask before a project.

None of this changes labelling. Our case pages say which voice is generated and which comes from people, whatever tool it was made with.

Our take and the limits

Voxtral 3 is a serious European offer for the whole speech path, and EU hosting takes the edge off a question that always comes up with real voices. But Mistral doesn't mention open weights for Voxtral 3 anywhere: in the model overview, all four models are listed as "Premier", while the older Voxtral Small from 2025 is marked "Apache 2.0". So self-deploying here means talking to Mistral, not simply downloading files.

What's still open is how the voices sound in an actual film: in Swedish, in Chinese, with emotion across a whole scene. And whether a clone from 30 minutes gets close to what we deliver today. We'll only know after our own test, and we haven't run it yet.

Sources

• Mistral AI: Voxtral 3, announcement of 20 October 2026
• Mistral Docs: Voxtral TTS 3, model page
• Mistral Docs: Voice cloning
• Mistral Docs: Languages
• Mistral Docs: Models, overview with licence labels
• Mistral AI: Usage Policy, effective 11 June 2026
• ElevenLabs Docs: Data residency
• ElevenLabs Docs: Professional Voice Cloning
• Miraki: Rapido
• Miraki: 21 ZUAN
• Miraki: When Everything Crumbles