Mistral releases Voxtral 3: speech recognition in 50+ languages, expressive TTS and a real-time voice-agent SDK

Mistral AI announced Voxtral 3 on 20 October 2026: a new set of speech models and a software kit for building voice agents that can hold real-time conversations and carry out tasks during a call.

What is in the release

• Voxtral Small Transcribe 3 — speech-to-text for 50 languages, including the 24 official EU languages, with speaker identification (diarization), context biasing with key terms, and end-of-turn detection. A smaller Mini version has 4 billion parameters.
• Voxtral Mini Transcribe Realtime 3 — real-time transcription that Mistral says runs at under 500 ms latency.
• Voxtral Mini TTS 3 — text-to-speech in 35+ languages and 50+ dialects, with voices that adjust emotion, pace and emphasis to context, plus “Pro Voice” adaptation from about 30 minutes of a speaker’s audio.
• A voice-agent SDK built on LiveKit that streams speech in, reasoning by a language model such as Mistral Small, tool calls and speech out; Mistral says each turn completes in under a second and works over phone (SIP) and web (WebRTC).

Prices and deployment

Mistral lists $0.0035 per minute for batch Mini transcription, $0.008/min for realtime, $13 per million characters for TTS and $0.04/min for the voice-agent API with Mistral Small 4. The models are available via Mistral’s EU, US and global API endpoints or for self-deployment, which the company pitches as a way to keep audio and retention under the customer’s control.

Caveat

Mistral says Voxtral 3 beats Whisper, ElevenLabs Scribe v2 and Gemini transcription models, but the charts are the company’s own evaluations, several judged by another AI model; independent benchmarks have not yet been published.