Microsoft's VibeVoice-ASR-Streaming: live "who said what" transcription, 7B open weights, 10 languages, no Turkish

Microsoft Research has released VibeVoice-ASR-Streaming, an open-weight streaming speech recognition model that transcribes as audio arrives and tags each segment with its speaker. Sizes: 1.5B and 7B, MIT license, on Hugging Face (whose counter shows 9 billion parameters for 7B, audio encoder included). The technical report reached arXiv 2 September, the GitHub announcement 3 September.

Per Microsoft, recognition and speaker diarization, traditionally separate tasks, are unified in one model. Product and person names passed as hotwords improve recognition. Ten languages: Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish. Turkish is not among them.

How it works

Audio is processed in chunks, text emitted once per chunk: 22 frames (2.9 seconds) or 15 frames (2.0 seconds), lookahead fixed at 0.5 seconds. The report puts expected speaker-attribution latency at 2.00 and 1.53 seconds. An NVIDIA GPU is required; hotwords go in via --context_info. Included: a WebSocket FastAPI demo and a file inference script; Transformers and vLLM work; recordings cap at 8 minutes.

The numbers

Across five multi-speaker meeting benchmarks, 7B averages 24.66 WER/CER and 31.55 speaker-attributed cpWER/cpCER. The report's table lists Gemini 3.5 Transcribe Live 25.23, GPT Realtime Whisper 39.31, GPT Live Transcribe 40.55, ElevenLabs Scribe v2 Realtime 41.39, Microsoft's own measurements, not an independent test. Against the offline VibeVoice-ASR, streaming costs 0.75 to 3.53 WER/CER points and 5.13 to 6.67 speaker-attribution points.

The family line

VibeVoice-TTS was open-sourced last August; the offline VibeVoice-ASR, which handles 60-minute recordings in one pass, followed on 21 January; ASR-BitNet, a compressed GPU-free version (4.62 GB down to 1.58 GB), came on 23 July. Streaming is the latest link.

For the corporate reader

Meeting minutes, call centers and voice command systems can now get speaker-separated live transcripts from an open model on-premises, data staying in-house. Without Turkish, organizations in Turkey cannot use it directly today; the open option remains the Whisper family. Microsoft warns the model “may produce unexpected, biased or inaccurate outputs” and advises against commercial or real-world use without further testing and development.

Sources

• Model card: huggingface.co/microsoft/VibeVoice-ASR-Streaming-7B
• GitHub, 3 September announcement: github.com/microsoft/VibeVoice
• Technical report: arXiv 2609.02812, 2 September: arxiv.org/abs/2609.02812