ModelCards
← All benchmarks

Artificial Analysis WER Index

We haven't written a description for this benchmark yet; see the card notes below for how each maker ran it.

Your workloadRAG over docs · 12K in · 600 out · 50K requests/mo · 30% cached
  1. 1Voxtral Small Transcribe 3 · Mistral AI
    2.1%
  2. 2MAI-Transcribe-2 · Microsoft
    2.0%

“Best value”: no other model here reports a higher score at a lower cost for your workload. Scores are self-reported by each maker with different settings, so treat close calls as ties.

How each maker reported it

  • Voxtral Small Transcribe 3 launch post (charts)

    Mistral's own evaluation. FLEURS: macro-average WER (CER for ja, ko, th, zh) over 45 languages supported by all seven systems compared, language specified (forced) or auto-detected (inferred). Diarization: macro-average cpWER over AMI-IHM, AMI-SDM, Switchboard, CallHome and AISHELL-4 (30-min versions of AMI and AISHELL). Keyphrase recall on Earnings-22 with context bias of key terms plus 0–100 distractors. Artificial Analysis WER Index as charted by Mistral. Lower is better except keyphrase recall.

  • MAI-Transcribe-2 p.2 (FLEURS chart), p.7 (Artificial Analysis chart)

    Microsoft's own comparison on the FLEURS test split, average WER over 60 languages, language forced or inferred. Footnotes: Gemini 3.5 Transcribe uses Gemini 3.1 Pro for Urdu; Gemini 3.6 Flash uses Gemini 3.1 Pro for Bulgarian (and Swahili inferred). Artificial Analysis WER Index (AgentTalk 50%, VoxPopuli 25%, Earnings-22 25%): MAI-Transcribe-2 2.0%, ranked #2 behind Fun-Realtime-ASR. Lower is better.

Only each model's own maker's number is shown; scores a maker reports for rival models are left out.