VibeVoice-ASR

Microsoft's MIT-licensed 8.7B speech-to-text model: an hour of audio in one pass with speaker labels, timestamps and 50+ languages.

In short. VibeVoice-ASR transcribes very long recordings, up to an hour, in one pass. Instead of plain text it tells you who spoke, when and what they said, and you can give it a list of hotwords such as names and technical terms. It covers more than 50 languages and is released under the MIT license.

Summary

Microsoft Research open-sourced VibeVoice-ASR on 21 January 2026 and described it in a technical report (arXiv 2601.18184). It is a unified speech-to-text model with about 8.7B parameters, stored as a 17.35 GB BF16 checkpoint. It accepts up to 60 minutes of continuous audio within a 64K-token context instead of cutting the audio into short chunks, so speaker tracking stays consistent across the whole recording. The model does transcription, speaker diarization and timestamping together, so its output says who said what and when. It accepts customized hotwords, supports more than 50 languages without a language setting, and handles code-switching.

Microsoft has kept extending the line. VibeVoice-ASR joined a Hugging Face Transformers release on 6 March 2026 and Azure AI Foundry Labs on 12 March. On 23 July 2026 Microsoft released VibeASR.cpp, an edge CPU inference engine that it says reaches real-time speed on three or more CPU threads with no GPU. On 3 September 2026 it added VibeVoice-ASR-Streaming, in 7B and 1.5B versions, which writes out who said what as speech arrives, supports hotwords, and covers 10 languages (Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish). The streaming report says its 7B model has the lowest average error rate across the five evaluation sets in that report.

The model cards show their diarization and error-rate comparisons only as images, so we do not quote figures here; read the technical reports before relying on any accuracy claim. Hugging Face showed about 714,000 downloads of VibeVoice-ASR in the 30 days to 11 November 2026, far more than for the streaming versions.

Good for

  • Transcribing long meetings, interviews and podcasts of up to an hour in one pass, with speaker labels and timestamps
  • Domain-specific audio where you can supply hotwords such as names and technical terms
  • Multilingual and code-switched speech in 50+ languages
  • Live, speaker-attributed captions with the streaming versions

Watch out for

  • The full model is large, about 17 GB in BF16, so it needs a big GPU unless you use Microsoft's CPU engine
  • The streaming versions cover 10 languages, fewer than the 50+ of the full model
  • The cards show results as charts only, so check the technical reports for actual numbers
  • All accuracy claims come from Microsoft's own reports

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of VibeVoice-ASR. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
ASR-Streaming-1.5B BF16 2.81B 5.63 GB 6.47 GB Laptop or phone
VibeVoice-ASR BF16 8.67B 17 GB 20 GB Gaming GPU
ASR-Streaming-7B BF16 8.67B 17 GB 20 GB Gaming GPU

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.

More about MIT

See also

Frequently asked questions

What is VibeVoice-ASR?
VibeVoice-ASR transcribes very long recordings, up to an hour, in one pass. Instead of plain text it tells you who spoke, when and what they said, and you can give it a list of hotwords such as names and technical terms. It covers more than 50 languages and is released under the MIT license.
Can I use VibeVoice-ASR commercially?
Yes. VibeVoice-ASR is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is VibeVoice-ASR open source?
VibeVoice-ASR is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run VibeVoice-ASR on my own computer?
Yes. The smallest variant, ASR-Streaming-1.5B, needs about 6.47 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was VibeVoice-ASR released?
VibeVoice-ASR was released by Microsoft on January 21, 2026.
How long a context window does VibeVoice-ASR have?
VibeVoice-ASR supports a context window of up to 64K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3-ASR Alibaba Qwen 918M – 2.35B – Apache 2.0 (Permissive) 2.12 GB Laptop or phone
Voxtral Realtime Mistral AI 4.43B 128K Apache 2.0 (Permissive) 10 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone