Mistral just dropped Voxtral 3, a new family of speech models plus a voice agent SDK that listens, thinks, calls tools and talks back in one real-time pipeline.
What’s in it
- Voxtral Small Transcribe 3: speech-to-text in 50+ languages with diarization (who said what), context biasing for technical terms, and end-of-turn detection. Mistral says it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe.
- Voxtral Mini Transcribe 3: a 4B model at $0.0035 a minute. The Realtime version runs under 500 ms latency.
- Voxtral Mini TTS 3: expressive speech in 35+ languages and 50+ dialects, with controls for emotion, pacing and emphasis, at $13 per million characters.
- Voice agent SDK: built on LiveKit. Transcribe, reason with a model like Mistral Small, call tools, then speak, with each turn under a second. It works over SIP phone lines and WebRTC in the browser.
- Pro Voice: voice adaptation from about 30 minutes of audio.
- Run it on Mistral’s EU, US or global API endpoints, or deploy it on your own infrastructure.
My take
The self-host option is the headline for me. Voice is about the most personal data there is, and being able to keep the audio on your own hardware and decide whether anything gets kept is a big deal. Each stage is a separate model, so you can swap the LLM or the voice without rebuilding the agent. That’s the same one-job-per-piece idea I like with bots. The cloning feature is powerful, so use it wisely.
Have a blessed day! And Happy Hacking!
