Mistral AI unveiled Voxtral 3 on October 20, 2026, a family of speech models plus a voice agent SDK designed to let AI agents listen, reason, speak, and act in real time. Transcription covers 50+ languages, text-to-speech spans 35+ languages and 50+ dialects with voices that adapt emotion and pace, and Pro Voice clones a speaker from 30 minutes of audio.
On October 20, 2026, Mistral AI unveiled Voxtral 3, a family of speech models and a voice agent SDK built for real-time, multi-step conversations. The suite pairs Voxtral Small Transcribe 3, a transcription model covering more than 50 languages — including the 24 official EU languages and many low-resource Asian and Indic languages — with Voxtral Mini TTS 3, a text-to-speech model spanning 35+ languages and 50+ dialects whose voices adjust their emotion, pace, and emphasis to the conversational context. Co-founder and chief scientist Guillaume Lample announced the release on X, describing the family as state of the art.
The transcription stack adds diarization, context biasing — where customer-supplied keyphrases improve accuracy on technical vocabulary without fine-tuning — and end-of-turn detection so agents respond without delay. Mistral reports the Small model achieves a 4.1% word error rate on FLEURS across 45 languages, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe. A cost-effective 4B-parameter Mini variant transcribes at $0.0035 per minute, while Voxtral Mini Transcribe Realtime 3 streams live transcription at roughly 240 milliseconds of latency, more accurate than GPT Live Transcribe at that speed. TTS is priced at $13 per million characters, which Mistral positions as high performance at a cost comparable to Gemini 3.8 Flash TTS.
For voice cloning, Voxtral's Pro Voice adaptation reproduces a speaker's distinctive character, emotional range, and speaking style across languages from 30 minutes of audio — a heavier requirement than Google's Gemini 3.8 Flash TTS, which clones from 30 seconds. The Voice Agent SDK, built on the open-source LiveKit real-time audio framework, combines listening, reasoning via an LLM such as Mistral Small, tool calling, and speaking into one streaming pipeline. A launch demo shows an agent reading a French doctor's referral letter aloud and booking a dermatologist appointment for an English-speaking patient.
The sovereign-AI angle is central to Mistral's pitch: Voxtral 3 runs through Mistral's EU-hosted API or can be self-deployed on customer infrastructure, giving organizations control over where audio is processed and whether conversations are retained. The launch follows Mistral's €3 billion Series D round in early September and comes seven months after its first Voxtral TTS model shipped with support for just nine languages. It also sharpens European competition with ElevenLabs, whose Eleven v4 claims more than 90 languages.
