Microsoft Open-Sources VibeVoice-ASR-Streaming-7B, Unifying Streaming ASR and Speaker Diarization
Microsoft has officially released and open-sourced VibeVoice-ASR-Streaming-7B on Hugging Face, a 7B-parameter LLM-based streaming speech recognition model with downloadable weights and an accompanying technical report (team includes Li Dong and Furu Wei). The model unifies speaker-attributed ASR and speaker diarization in a single end-to-end architecture, continuously transcribing "who said what" as audio arrives in real time rather than after full-audio processing.
Confirmed
- The model is named VibeVoice-ASR-Streaming-7B, 7B parameters, available on Hugging Face with open weights.
- A single model supports both streaming transcription and diarization, based on an LLM end-to-end architecture; it is among the first LLM-based end-to-end streaming speaker-attributed ASR systems.
- Covers 10 languages, with Chinese and English explicitly confirmed by Microsoft's official post.
- Supports custom hotwords to improve recognition of proper nouns and similar terms.
- It is the ASR edition of the VibeVoice series, targeting real-time transcription scenarios.
Why it matters
Traditional ASR pipelines typically run recognition and diarization as separate offline steps; unifying them in one streaming end-to-end model is more practical for latency-sensitive use cases such as live captions, meeting notes, and voice assistants. Open-sourcing a 7B model also lowers the barrier for local deployment and further development.
2026-09-02 ~ 2026-09-04 · 6 related posts
Primary sources
- Microsoft Releases VibeVoice-ASR-Streaming-7B for Streaming Chinese/English ASR — microsoft ·
- Microsoft releases VibeVoice-ASR-Streaming, unifying streaming ASR and speaker diarization — microsoft ·
- Microsoft open-sources VibeVoice-ASR-Streaming with real-time diarization in 10 languages — realmrfakename ·
- [source] Microsoft Releases VibeVoice-ASR-Streaming-7B for Streaming Chinese/English ASR — microsoft · 2026-09-02
- Microsoft releases VibeVoice-ASR-Streaming-7B open streaming speech recognition model — Acceptable-Cycle4645 · 2026-09-03
- Microsoft releases VibeVoice-Streaming ASR with real-time speaker attribution — donglixp · 2026-09-03
- [source] Microsoft open-sources VibeVoice-ASR-Streaming with real-time diarization in 10 languages — realmrfakename · 2026-09-03
- [source] Microsoft releases VibeVoice-ASR-Streaming, unifying streaming ASR and speaker diarization — microsoft · 2026-09-03
- Microsoft's VibeVoice-ASR-Streaming: first LLM-based streaming speaker-attributed ASR, 1.5B/7B open-sourced — realmrfakename · 2026-09-04