Skip to content- Sources: primary, discussion
- Summary: Voxtral 3 includes Small and Mini Transcribe 3 with diarization and context biasing in over 50 languages, a realtime transcription model under 500 ms latency, and Voxtral Speech 3 TTS at 13 dollars per 1M characters. The voice agent SDK is built on LiveKit and chains transcription, an LLM with tool calls, and speech into one pipeline over SIP or WebRTC. The models are on Mistral's API, and self-deployment requires contacting sales.
- Why it matters: Teams building voice agents get a full speech stack with per-minute pricing and a self-hosting option for data residency.