VibeVoice-ASR-Streaming

A streaming, speaker-attributed ASR model from Microsoft's VibeVoice line that transcribes who said what as audio arrives, with customizable hotwords for names and jargon and support for ten languages (Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish). The report presents it as one of the first LLM-based end-to-end approaches to streaming speaker-attributed recognition; earlier unified models such as VibeVoice-ASR merged transcription and diarization but worked offline, which the low-latency requirements of voice agents rule out. Released in 1.5B and 7B sizes under MIT, in the VibeVoice repository (54K GitHub stars).