ModelCards

Voxtral Small Transcribe 3

Audio modelCard summarized

Highest accuracy speech transcription model

Source: Voxtral 3 launch post and Voxtral audio docs (no model card published)

Use it via API

$0.004 / min
Code exampleCompare

Where to call it

  • Mistral API (Mistral Studio)voxtral-small-transcribe-3-0

Supports

No feature list in the maker's docs.

Full feature support

Good for

  • Transcribing recorded calls, meetings and podcasts
  • Diarization: who spoke when, with optional named speaker reference clips
  • Domain vocabulary via context biasing, without fine-tuning

Watch out for

Overview

Mistral AI · 03Transcribe, front

Small Transcribe

Highest accuracy speech transcription model

Typical price
$0.004 / min
Voxtral Small Transcribe 3 (batch)
Max length
—
Languages
50 languages
Modes
speech-to-text, speaker diarization
Released
20 Oct 2026

Voxtral Small Transcribe 3 is the larger of Mistral's two Voxtral 3 batch speech-to-text models. Mistral says it is state of the art in multilingual transcription and diarization, transcribes 50 languages (including the 24 official EU languages), labels speakers, and takes keyphrases or free-form text as context bias without fine-tuning. The smaller sibling, Voxtral Mini Transcribe 3 (4B parameters), offers the same offline features at $0.0035/min. launch post; Speech to Text docs

voxtral-small-transcribe-3-0
Typical price
$0.004 / min
Languages
50 languages
Weights
Closed
Released
20 Oct 2026
  • speech-to-text
  • speaker diarization
  • context biasing (keyphrases and free-form context)
  • segment timestamps
  • streamed response (SSE)