Mistral AI has introduced its Voxtral 3 suite of speech models, which debuted on October 20, 2026. The Voxtral Mini Transcribe 3 model, featuring 4 billion parameters, is specifically engineered for batch transcription tasks.
In recent evaluations by Voice Arena, the Mini model secured the second position among nine tested systems, achieving a Diarization Error Rate of 40.4%. It outperformed the larger Small Transcribe 3 model, which ranked fourth with a 41.1% error rate.
The Voice Arena leaderboard is based on 139 spontaneous English-language conversations verified by human reviewers. Both Voxtral models demonstrated an 83.5% accuracy rate in identifying the number of speakers present in a recording.
While the Mini model excels in speaker identification, the larger Small Transcribe 3 model maintains a slight advantage in word accuracy, scoring 2.1% on the Word Error Rate metric compared to 2.2% for the Mini.
The Voxtral series also showed improved performance in multilingual benchmarks, surpassing the Whisper large-v3 model on the FLEURS test across more than 45 languages. Despite these results, developers note that overlapping speech remains a significant challenge.
