Voxtral separates live speech from recorded audio

Mistral released Voxtral Transcribe 2, including Mini Transcribe V2 for recordings and Realtime for live applications. The announcement covers speaker separation and timestamps, while Realtime weights are available under Apache 2.0. Studio also gains an interface for trying transcription.

Context

Meeting notes can wait for a recording to finish; a conversational assistant cannot. Those uses balance delay and accuracy differently. Before adopting either approach, teams should test their own microphones, accents, background noise, and overlapping speakers. A broad benchmark rarely captures all the conditions of an actual conversation.