Guide · Open source
How to use VibeVoice: clone, set up and transcribe long audio
Updated · written with AI from the project's own documentation and checked against our catalogue data
VibeVoice is a family of open-source voice AI models from Microsoft that cover speech recognition and text-to-speech. This guide explains what the project includes, how to get the code, and which parts are suited to a developer self-hosting speech-to-text. By the end you will know where to start, what the README says about the models, and which limits to check before you build on it.
VibeVoicemicrosoft/VibeVoice · ★ 55.1k · MIT Install options & downloads Quick start
- 1Clone the repository. Run git clone https://github.com/microsoft/VibeVoice.git in a terminal to download the code.
- 2Read the model docs. Open docs/vibevoice-asr.md to see how the speech recognition model is set up and used.
- 3Pick your model. Choose VibeVoice-ASR for long recordings with speaker labels, or the streaming variant for live audio.
- 4Prepare hotwords. List names and domain terms from your audio to pass as customized hotwords.
- 5Test on a sample. Run a short recording first and check speaker labels and timestamps before processing long files.
What VibeVoice includes
VibeVoice is a family of models rather than a finished app. The main speech recognition model, VibeVoice-ASR, is designed to handle 60-minute long-form audio in a single pass and produce output that says who spoke, when, and what was said. It also supports customized hotwords, such as names or technical terms, to guide recognition.
The README also lists a streaming variant, VibeVoice-ASR-Streaming, which emits text once per audio chunk while recording is still in progress. A CPU-oriented BitNet variant is listed too, but its details are limited in this README.
- VibeVoice-ASR: long-form transcription with speaker and timestamp labels
- VibeVoice-ASR-Streaming: transcribes while audio is still arriving
- VibeVoice-ASR-BitNet: CPU variant, details limited in the README
- VibeVoice-Realtime-0.5B: real-time text-to-speech with streaming text input
Get the VibeVoice code
The install method is clone-and-build. Start by cloning the repository from GitHub. The README does not list prerequisites, so check the linked model documentation for each model before you set up an environment.
- Open a terminal in the folder where you keep projects.
- Run the clone command below.
- Read the docs folder, starting with docs/vibevoice-asr.md, for model-specific setup.
git clone https://github.com/microsoft/VibeVoice.gitChoose the right model for your task
Pick the model based on the job. For transcribing meetings or interviews with speaker labels, use VibeVoice-ASR. For live transcription, look at the streaming variant. Note that the README says the VibeVoice-TTS code was removed from the repository in September 2025, so do not plan a text-to-speech build around that model without checking the current state of the project.
The project homepage at microsoft.github.io/VibeVoice has demos and examples. The README also links a Playground for testing ASR without local setup.
- Transcription with speakers and timestamps: VibeVoice-ASR
- Live transcription: VibeVoice-ASR-Streaming
- Text-to-speech: check the README status first, since VibeVoice-TTS code was removed
Improve accuracy with hotwords and context
If your audio uses names, product terms, or jargon, the README says you can supply customized hotwords or background information. This is meant to guide recognition on domain-specific content. The README does not give a benchmark for how much this improves results, so test it on your own recordings.
- List names and technical terms that your audio contains.
- Compare transcripts with and without hotwords on a short sample.
- Check the finetuning-asr folder if you need to adapt the model further.
Understand the output and limits
VibeVoice-ASR jointly performs recognition, diarization (who spoke), and timestamping. Because it processes up to 60 minutes of continuous audio, it keeps speaker tracking across the whole recording instead of slicing it into short chunks. The README states the model supports over 50 languages, while the streaming variant lists 10 languages.
Most of these models are aimed at developers. Expect to handle GPU setup, model downloads from Hugging Face, and code integration yourself.
- Output includes speaker labels and timestamps
- Up to 60 minutes of audio in one pass for VibeVoice-ASR
- vLLM inference is mentioned in the README as a way to speed up inference
What people use it for
- Transcribing multi-hour meetings with speaker labels and timestamps
- Creating searchable transcripts of interviews without sending audio to a closed cloud service
- Improving recognition of product names and jargon with customized hotwords
- Live captioning with the streaming ASR variant
- Researching speech models using the published technical reports
Frequently asked questions
Is VibeVoice free to use?
VibeVoice is listed under the MIT licence in this directory entry, and the code is open source. Check the licence file in the repository and the license of each model weight on Hugging Face before commercial use.
What are the requirements for VibeVoice?
The README does not list prerequisites. The source notes that the BitNet CPU variant exists for low-resource devices, but other models are aimed at developers who can run them on suitable hardware. Check each model's documentation for details.
How do I install VibeVoice?
The install method is clone-and-build. Run git clone https://github.com/microsoft/VibeVoice.git, then follow the model documentation in the docs folder to build and run it.
Can VibeVoice do text-to-speech?
The README says the VibeVoice-TTS code was removed from the repository in September 2025. It still lists a Realtime-0.5B text-to-speech model with a Colab demo, so check the current README status before relying on TTS.
How does VibeVoice compare to Whisper?
Both are speech recognition options, but VibeVoice focuses on single-pass transcription of up to 60 minutes with speaker labels and timestamps. Whisper, WhisperX, pyannote-audio and Coqui TTS are comparable projects in this directory, so test each on your own audio.
Does VibeVoice work without a GPU?
The main models are not described as CPU-friendly in this README. The VibeVoice-ASR-BitNet variant is described as an edge CPU inference engine that needs no GPU, but the README gives limited details on it.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
VibeVoice suits developers who want to self-host speech-to-text on long recordings and need speaker labels and timestamps. It is not a good pick if you want a finished app with no setup, or if you need text-to-speech from the TTS model specifically, since the README says that code was removed.
Install options & downloads