Mistral AI’s Voxtral Mini Transcribe 3 takes second place on Voice Arena leaderboard

Mistral AI's Voxtral Mini Transcribe 3, a 4 billion parameter speech model, ranked second of nine models on Voice Arena's diarization and speech recognition leaderboard with a 40.4% Diarization Error Rate, edging out its larger sibling Small Transcribe 3 at 41.1% DER in fourth place. Voice Arena built the evaluation on 139 English-language spontaneous conversations with human-verified references, and both Voxtral models correctly identified speaker count 83.5% of the time. Mistral AI unveiled the Voxtral 3 suite on October 20, 2026, with Mini batch transcription priced at $0.0035 per minute versus $0.004/min for Small and $0.008/min for realtime, though Small held a slim word-accuracy edge on AA-WER at 2.1% versus Mini's 2.2%.

Mistral AI’s Voxtral Mini Transcribe 3 takes second place on Voice Arena leaderboard The 4 billion parameter speech model edged past its larger sibling in an independent test of who said what in real conversations Mistral AI https://cryptobriefing.com/markets/mistral/ ’s smallest new transcription model just beat its bigger sibling at one of the hardest jobs in speech recognition. Voxtral Mini Transcribe 3 ranked second out of nine models on Voice Arena’s diarization and speech recognition leaderboard. The larger Small Transcribe 3 landed in fourth. In AI, size usually wins the argument. This time, the little one got the last word. What Voice Arena actually measured Diarization is the part of transcription that figures out who is talking and when. Think of a court stenographer who not only types every word but also labels each line with the right speaker’s name. Voice Arena scores this with a metric called Diarization Error Rate, or DER. Lower is better. It roughly tracks how much of a conversation gets attributed to the wrong speaker, missed, or wrongly flagged as speech. Voxtral Mini Transcribe 3 posted a 40.4% DER , good for the number two spot among the nine models tested. Small Transcribe 3 came in at 41.1% DER , placing fourth on the same board. Voice Arena built its evaluation on real-world audio from 139 English-language conversations. These were spontaneous exchanges, not scripted reads, and the references were verified by humans. Mini performed particularly well on in-person recordings. Both Voxtral models correctly identified how many speakers were in a conversation 83.5% of the time. Overlapping speech remains the sticking point. When two people talk at once, the models still struggle to untangle them, and that is flagged as a key focus area for developers. The specs and the price tag Mistral AI unveiled the Voxtral 3 suite of speech models on October 20, 2026. Voxtral Mini Transcribe 3 is the flagship, built with 4 billion parameters and designed for batch transcription. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. The model supports a 32,000 token context. Pricing is where things get interesting for anyone running transcription at volume: Mini batch transcription: $0.0035 per minute Small batch transcription: $0.004/min Realtime transcription: $0.008/min So the cheaper model is also the one that ranked higher on diarization. Word accuracy tells a slightly different story Diarization is only half the picture. The other half is whether the model gets the words themselves right, which is measured by word error rate, or WER. On AA-WER, Mini scored 2.2%, ranking fifth. Small Transcribe 3 scored 2.1%, ranking third. So the bigger model holds a slim edge on raw word accuracy, while the smaller one does better at keeping speakers straight. On the multilingual side, benchmark results show the Voxtral models outscoring Whisper large-v3 on FLEURS WER across 45+ languages. FLEURS is a standard multilingual speech test, and Whisper large-v3 is a widely used reference point in the field. That matters because the Voice Arena diarization test was English only. What this means for the speech AI market A 40.4% DER still means a meaningful share of the conversation gets mislabeled, and overlapping speech remains an unsolved problem across the board. The test also covered 139 English conversations. That’s a rigorous sample of messy real-world audio, but performance in other languages, accents, or noisy environments may differ. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .