Voxtral Transcribe 2 ships with sub-200ms latency

Summary

Replaces chunked-audio adapters with native streaming architecture, enabling real-time voice agents without offline model hacks. Diarization and context biasing reduce post-processing overhead for meeting transcription and domain-specific vocabularies.

Why it matters

Replaces chunked-audio adapters with native streaming architecture, enabling real-time voice agents without offline model hacks. Diarization and context biasing reduce post-processing overhead for meeting transcription and domain-specific vocabularies.

Sources

1. 1.latency configurable down to sub-200ms
2. 2.approximately 4% word error rate on FLEURS and $0.003/min
3. 3.At 480ms delay, it stays within 1-2% word error rate, enabling voice agents with near-offline accuracy
4. 4.4B parameter footprint, it runs efficiently on edge devices
5. 5.Voxtral Realtime ships under Apache 2.0