Voxtral 3 separates listening, reasoning and speech in a voice agent
Mistral’s Voxtral 3 voice-agent SDK combines streaming transcription, a reasoning model that calls tools, and speech generation through LiveKit. The stages remain separate, allowing a developer to change the language model or voice without rewriting the whole agent. Speech recognition detects when a caller finishes, and the pipeline handles interruptions. Mistral also offers standalone speech generation; latency and language-quality claims need evaluation in the intended setting.
Mistral AI · Oct 20, 2026
Why it matters
A modular speech pipeline gives voice applications clearer control over transcription, business actions and spoken output.
More context
The release is available through Mistral’s paid API and Studio playgrounds. Standalone Voxtral Mini TTS 3 is listed at $13 per million characters and supports custom pronunciation. Try a preset voice with a short script; running the complete stack on private infrastructure requires contacting Mistral.
Keep in mind
Language coverage differs between transcription and speech generation. Published quality comparisons include vendor evaluations and model judges; neither end-to-end latency nor pronunciation quality was independently tested here.
Original read