Voxtral Mini Transcribe Realtime 3
Voxtral Mini Transcribe Realtime 3 processes live speech in a WebSocket session. It combines transcription with speaker diarization, context biasing and built-in end-of-turn detection. The October 2026 release is a public preview.
Preview
Tune when the assistant can respond
The model exposes a latency range and detects when a speaking turn ends. Context biasing lets the application supply relevant vocabulary. These features are documented together in one streaming session, rather than as separate offline processing stages.Voxtral Mini Transcribe Realtime 3 model page
Evaluate the audio workflow
Practical guidance. Measure how often the system cuts off a pause or waits after a completed sentence. The lowest latency setting may not give the best conversation. Include corrections, short answers and interruptions in a manual evaluation, and distinguish model latency from network and speech-generation time.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Preview |
| Launch date | |
| Access | API |
| Model identifiers | voxtral-mini-transcribe-realtime-3-0 |
| API identifier | voxtral-mini-transcribe-realtime-3-0Voxtral Mini Transcribe Realtime 3 model page |
| Configurable latency | 80 to 2,400 milliseconds, as documented by MistralVoxtral Mini Transcribe Realtime 3 model page |
| Access | Premier API preview; public weights not confirmedVoxtral Mini Transcribe Realtime 3 model page |
Related models
- Voxtral Mini Transcribe 3
Mistral AI · Voxtral
A preview speech-to-text model for recorded audio with timestamps and speaker diarization.
Launch date PreviewAPI - Voxtral TTS 3
Mistral AI · Voxtral
A preview speech-generation model with voice cloning, emotion tags and controls for speaking pace.
Launch date PreviewAPI