Mistral's Voxtral Mini Transcribe 3 ranks second on Voice Arena's diarization leaderboard
Mistral AI highlighted that its transcription model Voxtral Mini Transcribe 3 ranks second out of nine on a leaderboard that measures who spoke and what they said. According to Voice Arena, it even outperformed its larger version.

In AI model rankings, the biggest model usually wins. This time, according to data Mistral AI shared on X, a "mini" model managed to finish ahead of a larger one.
On October 28, the company posted that Voxtral Mini Transcribe 3 ranks second on Voice Arena's Diarization + ASR leaderboard. That leaderboard measures two tasks at once. ASR (automatic speech recognition) is converting audio into text. Diarization means identifying who is speaking at each point in a conversation, which is key for meetings, interviews and calls with multiple people.
The original figure comes from Voice Arena, the account that maintains the leaderboard. Six days earlier, on October 22, it had reported that the model placed second out of nine by DER, the diarization error rate, a metric where a lower value indicates fewer errors in attributing speaker turns. Voice Arena added that the model outperformed its "larger counterpart."
The evaluation came after the launch of Mistral's new Voxtral 3 family. According to Voice Arena, its Mini and Small transcription models were put to the test on that platform. The available text cuts off there, so it is not possible to confirm exactly which model was outperformed or the details of the rest of the evaluation.
The result should be read with caution. It is a position on a specific leaderboard, with nine models and a single metric, and the post Mistral shared does not include the DER values, the audio datasets used, the languages evaluated or the competing models. Nor is there information on pricing, licensing, availability by country or support for Spanish, details that matter for anyone who wants to use the tool in Latin America.
Even so, the signal is interesting for teams working with meeting recordings, contact centers or journalism, where knowing who said what is as important as the transcription itself. Before drawing conclusions, the sensible step is to review the Voxtral 3 launch, its technical documentation and its terms of use, and to test the model with your own audio, in Spanish and with regional accents.
Source: Mistral AI on X