VibeVoice-ASR

VibeVoice-ASR

An 8.7B speech-to-text model from Microsoft Research that transcribes up to 60 minutes in one pass, saying who spoke, when and what, with optional hotwords and support for over 50 languages.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VibeVoice-ASR is a speech-to-text model from Microsoft Research with 8.7B parameters. It takes up to 60 minutes of continuous audio in a single pass and returns a structured transcript that says who spoke, when, and what was said. You can supply hotwords, such as names or technical terms, to steer recognition. It supports over 50 languages, needs no language setting, and copes with speakers who switch languages.

Who it suits: It suits teams who transcribe long meetings, interviews, podcasts and lectures and need speaker labels and timestamps, and developers who build transcription tools.

Industry | Who uses it | How they use it | Example
Media and podcasts | Producers | Transcribe a whole episode with speaker labels | Feed in an hour of an interview and read who said what, with times
Business | Meeting-notes teams | Turn long recordings into structured notes | Add company and product names as hotwords to help recognition
Research | Speech researchers | Study multilingual speech with speaker and timing data | Run audio where speakers switch languages mid-sentence

If your audio is short or you need a small model, a lighter one may suit you better, since this one has 8.7B parameters.

Where to use it:

• Python projects that use the transformers library, as the Hugging Face page lists
• A computer with enough memory for the weights, about 17.3 GB in BF16
• Long recordings of up to 60 minutes at a time
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Meeting transcripts | Operations teams | Text with speakers and timestamps
Interview and podcast notes | Producers | A transcript of who, when and what for an hour of audio
Domain terms | Specialist teams | Better recognition of the names and technical words you supply

How to set it up: What you need first:

• Python and the Hugging Face command line
• Room for about 17.3 GB of BF16 weights, or about 4.3 GB at 4-bit, which is an estimate
• The installation steps on the project's GitHub README, which the card points to

1. Install the Hugging Face tools.
2. Download the model files.
3. Follow the installation and usage page that the card points to, since the card itself holds no code.

pip install -U huggingface_hub
hf download microsoft/VibeVoice-ASR

How to use it: The card holds no code of its own. It sends you to the GitHub README for installation and usage, and it also lists a demo, a fine-tuning guide and a vLLM guide. You give the model audio of up to 60 minutes and, if you like, a list of hotwords. It answers with a transcript that says who spoke, when, and what.

Part | What it does
Hotwords | Names, technical terms or background info that steer recognition
Single pass | Handles up to 60 minutes of audio at once, within 64K tokens
Rich transcription | Combines speech recognition, speaker labels and timestamps in one output

Example: You give it an hour-long interview and a few hotwords, such as the guest's name and a product name. You get a transcript that marks each speaker, the time of each turn and the words spoken.

For newcomers and experts: A good first try is a short recording of a few minutes with two speakers. See how the speakers and times are listed before you try an hour-long file. Get permission from the people in a recording before you transcribe it.

For experts, many speech models cut audio into short chunks and lose the wider context. This one accepts up to 60 minutes within 64K tokens, so speaker tracking and meaning stay consistent across the hour. One model does recognition, speaker labels and timestamps together instead of a chain of separate tools. The card shows charts for DER, cpWER and tcpWER, and gives the numbers only as images.

Up to 60 minutes of audio goes in
Hotwords you supply guide recognition
One model transcribes and labels speakers and times
You get who, when and what

Good to know:

Title: Model facts
Item | Value
Parameters | 8.7B
Weights in BF16 | about 17.3 GB
Weights at 4-bit | about 4.3 GB, an estimate
Longest audio | 60 minutes in one pass
Context | 64K tokens
Languages | over 50

• Evaluation: the card shows charts for DER, cpWER and tcpWER, so check the model page for the figures.
• Languages: over 50, with no language setting needed. The Hugging Face record names English, Chinese, Spanish, Portuguese, German, Japanese, Korean and French.
• Consent: get permission before you transcribe other people's voices.
• Feedback: the team asks for reports of unexpected or harmful behavior and says it will update the repository with fixes.
• License: MIT
• Cost to run: no API key or paid service named. You need hardware for an 8.7B model.
• Data it reaches: it downloads files from Hugging Face. If you run the model yourself, your audio stays on your machine.
• Not for: none named.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 8.7B
• Weights: about 17.3 GB in BF16; about 4.3 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 715,654
• Likes: 1,328
• License: mit
• Languages: en, zh, es, pt, de, ja, ko, fr
• Quantized versions: 12
• Fine-tunes: 20
• Tags: transformers, safetensors, vibevoice, transcriptoin, diarization, speech-to-text, automatic-speech-recognition

Links

• Model: https://huggingface.co/microsoft/VibeVoice-ASR
• Files: https://huggingface.co/microsoft/VibeVoice-ASR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/VibeVoice-ASR

Then follow the usage in its model card. Nothing in this kit runs by itself.