Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
The release adds multilingual speech-to-text and text-to-speech models covering over 50 languages alongside a streaming voice agent SDK built on LiveKit.
- Voxtral Small Transcribe 3 costs $0.004 per minute, while the 4B Mini variant costs $0.0035 per minute batch and $0.008 per minute in real time.
- Voxtral Speech 3 text-to-speech costs $13 per 1M characters, adapting emotion and pacing across 35+ languages and 50+ dialects.
- Features include speaker diarization, context biasing for domain vocabulary, acoustic and semantic end-of-turn detection, and voice cloning from 30 minutes of audio.
- The SDK supports real-time function calling over SIP and WebRTC, available via Mistral's hosted API or for self-deployment.
Developers building voice agents get a modular, low-latency streaming pipeline that can be hosted on sovereign EU endpoints or private infrastructure.
