Mistral launches Voxtral 3, its multilingual voice AI

Mistral has launched Voxtral 3, a voice AI platform that can listen, interpret, respond and take action in real-time conversations. The system combines speech transcription and synthesis models with an SDK for building agents that can handle calls, query systems and complete tasks in more than 50 languages.

The difference is that it does more than convert speech to text. Voxtral 3 coordinates the entire process: it detects when you have finished speaking, understands the request, queries external tools and responds with a voice that adapts its pace, emphasis and emotion to the context.

A conversation that can also get things done

Built on LiveKit, the Voxtral 3 SDK connects several models in a continuous workflow. One listens and transcribes, another reasons and uses tools, and a third generates the spoken response. According to Mistral, each turn can be completed in less than one second, and the agent stops speaking if you interrupt it.

In practice, an agent could verify your identity during a call, check the status of an invoice and book an appointment without human intervention. It can also connect to proprietary systems through function calls, web searches and internal information retrieval.

The technology works on phone calls through SIP, in browsers and apps through WebRTC, and with existing telephony platforms, phone systems and customer service centers.

Transcription in more than 50 languages

Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 transcribe conversations in more than 50 languages. In addition to recognizing the words, they can identify who said each sentence, a function known as diarization, and use a list of terms to improve accuracy with technical or business vocabulary.

They also include turn-end detection. The system combines acoustic signals with the meaning of the conversation to estimate when you have finished speaking, with the goal of responding quickly without cutting you off too soon.

The Voxtral Mini Transcribe Realtime 3 version is designed for voice agents and offers latency below 500 milliseconds in more than 35 languages. Its published prices are:

• Voxtral Mini Transcribe 3: $0.0035 per minute.
• Voxtral Mini Transcribe Realtime 3: $0.008 per minute.
• Voxtral Small Transcribe 3: $0.004 per minute.

Mistral says the Small model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe in its multilingual transcription and diarization evaluations.

More expressive, adaptable voices

Voxtral Speech 3 generates speech in more than 35 languages and 50 dialects. It can modify emotion, speed and emphasis based on the content, which is useful when a response needs to sound reassuring, urgent, formal or personable.

The system can also follow instructions about how to speak and adapt a voice from more than 30 minutes of audio. The Pro Voice feature aims to preserve not only the speaker's tone, but also their style and emotional range across different languages.

The announced price for voice generation is $13 per million characters. Mistral says this is cheaper than several alternatives from OpenAI, Deepgram, ElevenLabs and Google, although the final cost will depend on the volume and design of each application.

Control over the data

Voxtral 3 is available through Mistral's API, with European, US and global endpoints, and can also be deployed on your own servers or in a private cloud. This lets you decide where audio is processed and whether conversations are retained.

For a company, that control can matter as much as voice quality. A customer service center can automate part of its calls without necessarily sending every recording to external infrastructure.

What to watch now is how these agents perform outside demos: with noise, accents, interruptions and ambiguous requests. Voxtral 3 brings conversational voice closer to real tasks, but its usefulness will depend on how well companies connect the system to their data, rules and processes.