Mistral’s Voxtral 3: transcription, speech synthesis and an SDK for voice agents
The new generation of voice models covers over 50 languages for transcription and 35+ for synthesis. A voice agent starts at $0.04 a minute.
Mistral has released Voxtral 3: a set of speech recognition and synthesis models, plus an SDK that assembles them into a voice agent.
What’s new
- Voxtral Small Transcribe 3: the company’s most accurate transcription model, $0.004 a minute.
- Voxtral Mini Transcribe 3: batch transcription, $0.0035 a minute.
- Voxtral Mini Transcribe Realtime 3: streaming transcription with under 500 ms latency, $0.008 a minute.
- Voxtral Mini TTS 3: text-to-speech, $13 per million characters.
Transcription supports over 50 languages, including 24 official EU languages, along with speaker diarization and custom vocabulary. Synthesis covers 35+ languages and offers voice cloning.
Voice agents
The SDK is built on LiveKit and joins speech-in, reasoning and speech-out into one stream, with SIP and WebRTC support. Each stage is a separate model, so you can swap the LLM, voice or language without rewriting the agent. The voice agent API costs $0.04 a minute with Mistral Small 4, or $0.08 with a larger model.
How to try it
The models are available through Mistral’s API with European, US and global endpoints, and you can test them in Studio. For self-deployment, you need to contact the company.
Limitations
The announcement doesn’t name a license. It lists both 35+ and 40+ languages for the streaming model, so check support for your language yourself.
Prepared with AI from public sources.