Mistral has released Voxtral 3, a speech model family and voice agent toolkit built for real time conversations. The set includes Voxtral Small Transcribe 3 for speech to text, Voxtral Mini TTS 3 for speech generation, and a voice agent SDK that chains listening, reasoning, tool calling and speaking into a single streaming pipeline. Transcription now covers 50 or more languages, among them all 24 official EU languages plus several low resource Asian and Indic languages. The speech model supports 35 or more languages and over 50 dialects, with voices that shift emotion, pace and emphasis to match the conversation.
Mistral positions the transcription model as state of the art on multilingual accuracy and speaker diarization, the task of working out who said what. It adds context biasing, where keyphrases and domain vocabulary are supplied at request time to lift recall on technical terms without fine tuning, plus end of turn detection so an agent can respond without waiting. Mistral says Small Transcribe 3 outperforms Whisper large-v3, Scribe v2 and Gemini 3.5 Transcribe, and that the Mini variant runs at 4B parameters and costs about USD 0.0035 per minute of audio. A Pro Voice tier reproduces a speaker's distinctive character and emotional range from 30 minutes of audio.
Deployment is much of the pitch. Voxtral 3 is available through Mistral's EU hosted API or for self deployment on customer infrastructure, which lets organisations decide where audio is processed and whether conversations are retained. Mistral frames the target market as contact centres, service desks, virtual assistants and hands free copilots, where a voice agent has to read intent, pull records from business systems and finish multi step tasks mid call. The release lands in a busy stretch for the company, arriving a day after Mistral and CERN signed a statement of intent to build a Large Physics Model.