Model · Microsoft · Speech-To-Text
VibeVoice-ASR
Microsoft's open long-form transcription model. It takes up to an hour of audio in one pass and returns who spoke, when, and what they said — transcription, speaker labels and timestamps together, in 50+ languages.
- Modality
- Speech-To-Text
- License
- MIT (Open)
- Parameter size
- 8.7B
- Released
- January 21, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
What you need to run it
- Mac
- Yes, but slow. Any Apple Silicon Mac (M1 or newer) with at least 24 GB (32 GB recommended) of memory. It runs, but much slower than on an NVIDIA card.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 24 GB of video memory.
- No graphics card
- Works. Runs on a normal computer's processor, just slower.
The full model wants a 24 GB graphics card or Mac; the 1.58 GB BitNet version runs in real time on an ordinary CPU, Apple M4 included, with a small accuracy drop.
Numbers are for: full 8.7B in bf16; a separate 1.58 GB BitNet build (VibeASR.cpp) runs on a CPU. Our estimate from the model’s size — the maker publishes no figure.
Strengths
- MIT licence — free for commercial use
- 60 minutes of audio in a single pass, so speaker tracking stays consistent across a whole meeting
- Speaker diarization and timestamps built in — no separate diarization tool
- Custom hotwords (names, jargon) to improve domain accuracy
- 50+ languages
- A 1.58 GB BitNet build (VibeASR.cpp) runs in real time on a CPU with 3+ threads — no GPU needed
- Streaming 7B variant (2026-09-03) transcribes who-said-what as audio arrives, in 10 languages
Weaknesses
- The full 8.7B model is a GPU workload; the CPU path needs the separate BitNet build
- Microsoft pulled the sister VibeVoice-TTS code in 2025-09 after misuse — a reminder the project can withdraw components
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | MIT; Transformers version at microsoft/VibeVoice-ASR-HF |
| GitHub | local | Reference code, vLLM serving and fine-tuning |
| VibeASR.cpp | local | CPU inference engine for the BitNet build |
| Azure AI Foundry Labs | hosted | Try it in Microsoft Foundry |
Official sources
- Model card model
- GitHub github
- Technical report paper
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: https://github.com/microsoft/VibeASR.cpp + size.
- — Initial entry. Added 2026-10-28 from the popularity sweep: 55k GitHub stars (whole VibeVoice repo), 677k Hugging Face downloads/month. Open-sourced 2026-01-21 per the README news; BitNet CPU build 2026-07-23; streaming 7B model 2026-09-03. The README also records Microsoft removing the VibeVoice-TTS code on 2025-09-05 over misuse.
Esc