Mistral Releases Voxtral 3 Real-Time Voice Agent SDK for 50+ Languages

Mistral has released Voxtral 3, a family of speech-to-text and text-to-speech models along with a voice agent software development kit designed to support real-time, multi-step conversations across more than 50 languages. The company says the system combines accurate transcription, natural and emotionally adaptive speech synthesis, reasoning and tool calling in a single streaming pipeline, enabling voice agents to listen, interpret intent, access connected systems and respond during live calls or sessions. The release includes Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3 for offline transcription, with support for diarization, context biasing and end-of-turn detection, as well as a realtime transcription model that achieves sub-500-millisecond latency in more than 35 languages. For speech generation, Voxtral Mini TTS 3 and Voxtral Speech 3 support 35 or more languages and 50 dialects, adjusting emotion, pacing and emphasis based on conversation context. Mistral also offers Pro Voice adaptation, which can reproduce a speaker’s character and style from roughly 30 minutes of audio. The voice agent SDK, built on LiveKit, connects listening, reasoning, function calling and speech generation so each conversational turn can complete in under one second. It can integrate with phone systems, browsers, apps, IVR platforms and contact-center software through SIP, WebRTC and function calling, while allowing enterprises to change the language model, voice or language without rebuilding the agent. Voxtral 3 is available through Mistral’s API on European, US and global endpoints, with EU-hosted deployment emphasizing data sovereignty, and can also be self-deployed. The company positions the release for contact centers, virtual assistants, service desks and hands-free copilots requiring low-latency, multilingual and controlled voice automation.