Mistral AI officially released the new Voxtral3 voice model series and voice agent SDK on October 21, 2026. The product lineup includes the Voxtral Small Transcribe3 transcription model, the Voxtral Mini TTS3.0 speech synthesis model, and a voice agent SDK built on LiveKit. It supports access via API or self-hosted deployment on private clouds and your own infrastructure, and is designed to provide high-performance real-time streaming voice interaction capabilities for call centers, service desks, virtual assistants, and hands-free driving systems.
In terms of technical performance, Voxtral Small Transcribe3 supports more than 50 languages and offers speech segmentation, context biasing, and turn-end detection. In the FLEURS chart, its forced error rate and inference error rate are 4.1% and 4.0%, on par with Microsoft AI Transcribe2 and better than Gemini3.5Transcribe, Scribe v2, GPT Transcribe, and Whisper large-v3.
The 4-billion-parameter Voxtral Mini Transcribe3 costs $0.0035 per minute for batch use and $0.008 per minute for the real-time version, with latency under 500 milliseconds. Meanwhile, Voxtral Mini TTS3.0 provides speech synthesis in over 35 languages and more than 50 dialects, priced at $13 per million characters. It supports tone adaptation, prompt instructions, and Pro Voice cross-lingual voice cloning based on 30 minutes of audio.
The SDK built on LiveKit integrates transcription, reasoning, tool calling, and speech generation into the same real-time streaming pipeline, enabling general-purpose models to reason and call tools autonomously while handling conversations. It keeps conversation turns under one second and automatically stops the agent when the caller interrupts. Analysts believe Mistral's release balances model performance with enterprise-grade deployment control and, through a modular architecture and private deployment support, further accelerates the large-scale deployment of voice AI agents in complex business scenarios.