Mistral releases Voxtral 3: connecting speech recognition, reasoning, and tool calling into a real-time Agent

Beating AI News Flash: French AI company Mistral has released Voxtral 3, including three speech transcription models, one speech synthesis model, and a complete Voice Agent SDK.

In the official demo, a user asks in English to book a dermatologist appointment. The AI reads a French referral letter, calls the system to check available time slots, completes the booking, and then confirms in English. The entire process is conducted through real-time voice conversation.

Voxtral 3 is not an end-to-end speech large model that handles all tasks. It adopts a modular design: the transcription model is responsible for recognizing speech and determining when the user has finished speaking, general large models such as Mistral Small handle reasoning and tool invocation, and the speech synthesis model handles responses. The SDK based on LiveKit strings these components into a real-time pipeline, supports phone and web calls, and also allows users to interrupt at any time. Mistral claims that single-turn interaction latency is less than 1 second. Developers can also replace the reasoning model or speech model independently.

Transcription-supported languages have expanded from 13 in the previous generation to more than 50, and speech synthesis supports more than 35 languages and more than 50 dialects. In Artificial Analysis's independent non-real-time transcription test, Voxtral Small Transcribe 3 had a word error rate of 2.1%, ranking third, behind only models from Alibaba and Microsoft. Speech synthesis ranked 13th.

The API charges $0.008 per minute for real-time transcription and $13 per million characters for speech synthesis. Enterprises can also contact Mistral to deploy on their own servers or private cloud.