Voxtral Small Transcribe 3 logo

Voxtral Small Transcribe 3

mistral · Released Oct 2026

Conditional

Conditional. Voxtral Small Transcribe 3 is Mistral's flagship speech-to-text model, billed per minute on the Mistral API with diarization and context biasing, and Mistral's own FLEURS and diarization charts put it ahead of Whisper large-v3, GPT Transcribe and Scribe v2, tied with Microsoft AI Transcribe 2 on FLEURS. The deciding caveat is that those charts are Mistral's, the model docs label it Public Preview, and no open weights are stated. Suits teams that want regional-endpoint transcription with speaker labels and can validate accuracy on their own audio.

Is it right for you?

Good for

  • Meeting, call and interview transcription that needs speaker labels; Mistral lists diarization and a diarization cpWER chart (18.4% for Small vs 18.6% for Mini, Mistral-reported)
  • Domain vocabulary such as earnings calls, where context biasing is documented (Mistral reports Earnings 22 keyphrase recall of 81.4% without context and 94.6% with context; Scribe v2 is 84.0% without and 94.3% with context, so Small does not lead without context and leads by only 0.3 points with it)
  • Multilingual transcription; the launch post says 50+ languages and Mistral's FLEURS average over the 45-language intersection is 4.1% WER with the language specified, 4.0% auto-detected
  • Buyers who want regional endpoints; the post says the API is hosted on European, US and global endpoints, with on-premises or private cloud by talking to Mistral

Not good for

  • Self-hosting from open weights; the pages we read state no open weights, the model is listed as Premier tier, and the launch post routes full on-premises or private cloud deployment through Mistral's team
  • Live, low-latency streaming transcription; that is the separate Mini Transcribe Realtime 3 model
  • Anything that needs accuracy claims confirmed outside Mistral: every benchmark we read is Mistral's own chart

How it performs by task

Transcription accuracy

Very Good

Mistral reports 4.0 to 4.1% WER on its FLEURS average over 45 languages, ahead of Whisper large-v3, GPT Transcribe, Scribe v2 and Gemini 3.5 Transcribe and tied with Microsoft AI Transcribe 2; on Mistral's Artificial Analysis WER Index chart Microsoft AI Transcribe 2 is ahead of Small; vendor charts, untested by us

Speaker diarization

Very Good

Mistral reports 18.4% cpWER across AMI, Switchboard and AISHELL-4, against 19.0% for Pyannote Orchestrator and 20.3% for Microsoft AI Transcribe 2; vendor chart, untested by us

Domain vocabulary

Good

Context biasing is documented and Mistral publishes a keyphrase recall chart; untested by us

Pricing

Input

$0.004 / min of audio

Output

Billed per minute (single rate)

Context

Per-minute billing; 32k context per docs

View full pricing

Benchmarks

BenchmarkScoreSource
FLEURS 45-language macro WER, language specified (Mistral-reported)4.1% Source
FLEURS 45-language macro WER, language inferred (Mistral-reported)4.0% Source
Diarization cpWER (Mistral-reported)18.4% Source

No verdict changes yet

The clock starts day one. Changes land here as our verdict evolves.

Verification log

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

How we evaluate