Voxtral Mini Transcribe Realtime 3
Audio modelCard summarizedA realtime transcription model built for agentic use-cases
Source: Voxtral 3 launch post and Voxtral realtime transcription docs (no model card published)
Where to call it
- Mistral API (Mistral Studio)
voxtral-mini-transcribe-realtime-3-0
Good for
- Voice agents that must know when the user has finished speaking
- Live captions and real-time note taking
- Phone lines and microphones, with live speaker labels
Overview
Mistral AI · 01Realtime, front
Mini Realtime 3
A realtime transcription model built for agentic use-cases
Typical price
$0.008 / min
Realtime transcription
- Max length
- —
- Languages
- 45 languages and dialects
- Modes
- speech-to-text, realtime streaming (WebSocket)
- Released
- 20 Oct 2026
Voxtral Mini Transcribe Realtime 3 is Mistral's streaming speech-to-text model for voice agents and live captions. It runs one WebSocket session per stream, writes text while people speak with a streaming delay set per session (80 ms to 2.4 s), and adds end-of-turn detection that Mistral says uses both acoustic and semantic information, plus diarization and context biasing. It is the listening stage of Mistral's new voice agent SDK. launch post; Realtime transcription docs
voxtral-mini-transcribe-realtime-3-0- Typical price
- $0.008 / min
- Languages
- 45 languages and dialects (docs); launch post says 35+ / 40+
- Weights
- Closed
- Released
- 20 Oct 2026
- speech-to-text
- realtime streaming (WebSocket)
- configurable streaming delay (80–1,200 ms or 2,400 ms)
- end-of-turn detection
- speaker diarization
- context biasing