Mistral releases Voxtral 3 for building voice agents that take calls

Mistral AI, a French AI company, has released Voxtral 3. It is a set of tools for building voice agents: software that can listen to a caller, work out what they need, and speak back. It can also take action, such as booking an appointment. It is available now.

What it does

Mistral's demo shows the idea. A patient says her doctor gave her a referral letter in French and asks for a booking. The agent reads the letter, explains it, offers a Tuesday 5:45 pm slot with a dermatologist, and books it. It then says what to bring and offers to text the details in English.

Voxtral 3 has three main parts:

• Speech-to-text that covers 50+ languages and can tell who said what.
• Text-to-speech voices that adjust emotion and pace, in 35+ languages and 50+ dialects.
• A toolkit (SDK) that links listening, thinking, tool use and speaking in one live system, over phone or web.

Smoother conversations

The speech-to-text model can detect when someone has finished speaking, so the agent does not talk over people or leave awkward gaps. It also accepts a list of key terms, such as product names or medical words, to get them right without extra training. Mistral says its larger transcription model beats Whisper, Scribe v2 and Gemini 3.5 Transcribe in its own tests.

Cost and privacy

The smaller transcription model, Voxtral Mini Transcribe 3, costs $0.0035 per minute. Mistral says it has 4 billion parameters, a measure of model size, and is built for fast, accurate transcripts.

Businesses can use Mistral's EU-hosted service or run Voxtral 3 on their own servers. Either way, they choose where audio is processed and whether any recording is kept. That matters for clinics, law offices and anyone handling sensitive calls.

A separate feature, Pro Voice, can copy a speaker's style and emotional range across languages from just 30 minutes of audio.

Who it affects

Mistral points to contact centers, service desks, virtual assistants and hands-free helpers. For a small business, the pitch is that phone answering and booking by voice gets cheaper to build, and call audio does not have to leave your control. Callers may hear more of these agents, in more languages.

What is not clear yet

The material does not give prices for the larger transcription model, the voices or the SDK, so total costs are unknown. The performance claims come from Mistral, not outside testers. It is also not stated how well the agents handle noisy phone lines or unusual requests in real use.