What it is
VibeVoice-ASR is an open speech-recognition model from Microsoft Research, released under the MIT licence on January 21, 2026. It processes up to 60 minutes of audio in a single pass within a 64K-token window, and produces a structured transcript of who spoke, when and what they said, combining transcription, speaker diarisation and timestamps. It supports more than 50 languages, handles code-switching and accepts custom hotwords for names and jargon.
It belongs to the VibeVoice family, whose history matters. Microsoft released the VibeVoice text-to-speech model in August 2025, then removed its code from GitHub in September 2025 after finding it was 'used in ways inconsistent with the stated intent'. The TTS weights remain on Hugging Face, with an audible AI disclaimer and watermark built in, but are marked for research only. The ASR line has continued: a CPU-friendly BitNet build followed in July 2026 and a Streaming version in September 2026.
Our review
VibeVoice-ASR is the open model to reach for when the recording is a long meeting rather than a short clip. Whisper and Parakeet return words; working out who spoke when is a second model with its own errors. Microsoft's model does both over up to 60 minutes in one 64K-token window, which is what keeps speaker labels consistent from the first minute to the last.
Microsoft now publishes its numbers as text rather than only charts. On the multilingual MLC-Challenge meeting set it reports an average diarisation error rate of 3.42% and WER of 12.07% across 11 languages, from 7.99% in English to 21.54% in Portuguese. On far-field English from a single distant microphone (AMI-SDM), WER rises to 24.65%. These are self-reported meeting benchmarks, not the read-speech sets most leaderboards use, and we could not find the model on Artificial Analysis's speech-to-text board, where hosted Parakeet scores 4.5% and ElevenLabs Scribe v2 2.2%.
It is heavy for a transcriber. Hugging Face lists about 9 billion parameters and 17 GB of BF16 weights, against 0.6 billion for Parakeet, and Microsoft's install guide starts from an NVIDIA PyTorch container with flash attention. The BitNet build, VibeASR.cpp, shrinks the model to 1.58 GB and runs faster than real time on three CPU threads, by Microsoft's account. The MIT licence permits commercial use, yet the README calls the models 'intended for research and development purposes only' and advises against commercial use without further testing, a line legal teams will read closely.
The family's misuse history belongs to speech synthesis: Microsoft pulled the VibeVoice TTS code in September 2025. Transcription carries no impersonation risk, and the ASR line has kept shipping, with Transformers support and an Azure AI Foundry Labs listing in March 2026 and a streaming model with hotwords in September. That streaming model covers 10 languages, a fifth of the batch model's range.
Where it earns its place
A platform engineer at a 200-person firm with a spare data-centre GPU
Transcribe recorded meetings with speaker labels without sending audio out
Serve the model with vLLM and send each recording of up to an hour as one request, passing attendee names and project terms as hotwords. There is no per-minute fee; the cost is GPU time, and audio never leaves the network.
A podcast production company publishing 40 hours of interviews a month
Get speaker-labelled transcripts for show notes and captions
One pass per episode returns who said what with timestamps, replacing a Whisper-plus-diarisation pipeline. For comparison, 2,400 minutes on ElevenLabs Scribe v2 is about $9 a month at ElevenLabs' $0.22 an hour, so self-hosting pays only if the GPU exists already or the audio must stay in-house.
A field researcher recording interviews on a laptop
Transcribe without a GPU or a cloud service
VibeASR.cpp, the BitNet build, is 1.58 GB and runs faster than real time on three CPU threads according to Microsoft. Its accuracy against the full model is not published in the figures we read.
Before you commit
- Setup
- Microsoft's guide starts from the NVIDIA PyTorch 25.12 container, installs flash attention and the repo with pip; Transformers has loaded it natively since March 2026, and a separate guide covers vLLM serving.
- Integrations
- Transformers, vLLM, VibeASR.cpp for CPU and Azure AI Foundry Labs; no Hugging Face inference provider hosts it. Fine-tuning instructions are on GitHub.
- Data and privacy
- Self-hosted, so audio stays on your hardware. The MIT licence allows commercial use, but the README describes the models as for research and development and asks for further testing before commercial deployment.
- Cost at scale
- No hosted price exists. Self-hosting means about 17 GB of BF16 weights on a GPU, or the 1.58 GB BitNet build on CPU; for scale, Together AI charges $1.50 per 1,000 minutes for hosted Parakeet and ElevenLabs about $3.67 (its $0.22 an hour) for Scribe v2.
- Limits
- 60 minutes per pass in a 64K-token window, so longer recordings must be split. Self-reported WER runs from 7.99% on English meetings to 24.65% on far-field AMI-SDM; the streaming variant covers only 10 languages.
What's good
- Transcript, speaker labels and timestamps in one pass
- Up to 60 minutes per pass, 50+ languages, custom hotwords
- MIT licence; Transformers and vLLM support
- 1.58 GB BitNet build runs on CPU
What's not
- About 9B parameters and 17 GB of weights
- Only self-reported accuracy; 24.65% WER on far-field AMI-SDM
- README discourages commercial use without further testing
- No hosted inference provider; Azure access is through Foundry Labs
Best for
Teams transcribing long, multi-speaker recordings on their own hardware who want speaker labels without a separate diarisation step.
Alternatives
VibeVoice-ASR: common questions
- Is VibeVoice-ASR worth it?
- VibeVoice-ASR turns an hour-long meeting into a speaker-labelled transcript in one pass, and every accuracy figure for it is still Microsoft's own. We rate it 3.8 out of 5.
- How much does VibeVoice-ASR cost?
- Free under the MIT licence. Also available in Azure AI Foundry Labs.
- What are the downsides of VibeVoice-ASR?
- About 9B parameters and 17 GB of weights; only self-reported accuracy; 24.65% WER on far-field AMI-SDM; rEADME discourages commercial use without further testing; no hosted inference provider; Azure access is through Foundry Labs.
- Who is VibeVoice-ASR best for?
- Teams transcribing long, multi-speaker recordings on their own hardware who want speaker labels without a separate diarisation step.
- How much does VibeVoice-ASR cost for a team or in production?
- No hosted price exists. Self-hosting means about 17 GB of BF16 weights on a GPU, or the 1.58 GB BitNet build on CPU; for scale, Together AI charges $1.50 per 1,000 minutes for hosted Parakeet and ElevenLabs about $3.67 (its $0.22 an hour) for Scribe v2.
- Is VibeVoice-ASR safe for company data?
- Self-hosted, so audio stays on your hardware. The MIT licence allows commercial use, but the README describes the models as for research and development and asks for further testing before commercial deployment.
- What are the best VibeVoice-ASR alternatives?
- The closest alternatives we have reviewed are Parakeet TDT 0.6B v3, Whisper large-v3-turbo and ElevenLabs Scribe v2.
Answers checked Oct 23, 2026.