Mistral launches Voxtral 3 speech models and voice agent SDK

Mistral has introduced Voxtral 3, a family of speech models and a voice agent SDK designed for live conversations and multi-step workflows. Built on LiveKit, the SDK connects speech recognition, a reasoning model, tool calling and voice generation in a streaming pipeline. Agents can retrieve information, verify callers, check bills and book appointments through connected business systems. Mistral says conversational turns complete in under one second, with agents stopping when callers interrupt.

The transcription models support 50+ languages, speaker identification and contextual vocabulary prompts to improve recognition of specialized terms without fine-tuning. The Mini batch model costs $0.0035/min, while realtime transcription costs $0.008/min. Speech generation supports 35+ languages and 50+ dialects, with controls for emotion, pacing, emphasis and pronunciation. Pro Voice adaptation can reproduce a speaker’s distinctive style across languages using 30 minutes of audio.

Organizations can use Mistral’s EU-hosted API or deploy the stack on their own infrastructure, controlling where audio is processed and whether conversations are retained. European, US and global API endpoints are also available. The modular architecture lets developers change the reasoning model, voice or language without rewriting the agent, while supporting telephone calls, browsers, apps and existing contact-center systems.