Voxtral Mini Transcribe 3 ranked second among nine models tested by Voice Arena, ahead of fourth-place Small Transcribe 3, with a diarization error rate of 40.4% versus 41.1%. The evaluation used 139 English-language conversations featuring spontaneous exchanges and human-verified references; both models correctly identified the number of speakers in 83.5% of cases. Overlapping speech nevertheless remains difficult to process. Small Transcribe 3 achieved a better word error rate of 2.1%, compared with 2.2% for Mini, while the Voxtral models outperformed Whisper large-v3 on the multilingual FLEURS benchmark covering more than 45 languages. Unveiled on October 20, 2026, the Voxtral 3 suite includes a 4-billion-parameter Mini model, a 32,000-token context window, and pricing of $0.0035 per minute for batch transcription, $0.004 for the Small version, and $0.008 for real-time transcription.
Source: Read the original article
