AI Research Atlas

Voxtral Transcribe 2 (Mini Transcribe V2 and Realtime)

Mistral AI · 4 February 2026

Voxtral Realtime, open under Apache 2.0, delivers streaming transcription with sub-200ms configurable delay; Mini Transcribe V2 hits about 4% WER on FLEURS.

Two speech-to-text models. Batch Mini Transcribe V2 has 13 languages, diarization, word timestamps and a $0.003/min price, and Realtime ($0.006/min) holds 1-2% WER at 480ms delay per Mistral and matches batch accuracy at 2.4 seconds.

Date
Wednesday, 4 February 2026
Lab
Mistral AI
Kind
model
Access
open weights
Price
$0.003/min (batch), $0.006/min (realtime), Feb 2026

Figures

MeasureValueMeasured by
FLEURS WER (Mini Transcribe V2)~4%company
Realtime latencysub-200ms configurable
480ms delay keeps WER at 1-2% per Mistral
company

Only the Realtime model is open weights; Mini Transcribe V2 is API.

Sources

  1. mistral.ai/news/voxtral-transcribe-2

This record was checked against its sources on 6 October 2026. How we check

Related