Voxtral Transcribe 2 splits batch transcription from live speech

Mistral released Voxtral Transcribe 2 on 4 February, pairing Mini Transcribe V2 for recorded audio with Voxtral Realtime for streaming applications. The realtime model ships under Apache 2.0 and supports configurable delay down to below 200 milliseconds. Both cover thirteen languages, including Hindi.

The batch model adds speaker labels, word timestamps and vocabulary guidance. Mistral notes that overlapping speech is typically transcribed as one speaker, while context biasing is optimised for English. Those details matter for meetings and interviews, where names and interruptions can carry meaning. A lower streaming delay may suit a voice interface; a recorded transcript can wait for more context. Developers should choose the endpoint and review process for that difference rather than treating transcription as one uniform task.