# VibeVoice — card for AI agents

Long-form multi-speaker TTS for podcast-style dialogue, up to 90 minutes and 4 speakers, in English and Chinese.

Category: Voiceover (text-to-speech) · maker Microsoft Research · checked 2026-10-25

## Price
- Price: Free
- Free tier: Open weights, free to download.

## License
MIT (model card limits intended use to research) — https://huggingface.co/microsoft/VibeVoice-1.5B

- commercial: **conditions** (Commercial use with conditions)
  > The VibeVoice model is limited to research purpose use exploring highly realistic audio dialogue generation
  (https://huggingface.co/microsoft/VibeVoice-1.5B)
- attribution: **required** (Credit required)
  > The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
  (https://github.com/microsoft/VibeVoice/blob/main/LICENSE)
- in app: **conditions** (Ship inside your app with conditions)
  > We do not recommend using VibeVoice in commercial or real-world applications without further testing and development.
  (https://huggingface.co/microsoft/VibeVoice-1.5B)

## Rules to follow
- MIT-licensed weights, but Microsoft states the model is for research use and does not recommend commercial use.
- Every generated file carries an audible AI disclaimer and an imperceptible watermark.
- Not licensed for: voice cloning without recorded consent, disinformation, live deepfake voice conversion.
- Microsoft removed the TTS inference code from GitHub in September 2025; the 1.5B weights stay on Hugging Face.

## How to connect
- Download only
- Repository: https://github.com/microsoft/VibeVoice
- Website: https://huggingface.co/microsoft/VibeVoice-1.5B

## Notes
Hugging Face API, Oct 25, 2026: 324867 downloads, 2506 likes; GitHub microsoft/VibeVoice: 54964 stars.

## Numbers (AgentGid data)
- Gid Score: 52
- GitHub stars: 57,295

## Lines to paste into your prompt or AGENTS.md

```
- VibeVoice (Voiceover (text-to-speech)). Long-form multi-speaker TTS for podcast-style dialogue, up to 90 minutes and 4 speakers, in English and Chinese. Price: free; commercial use with conditions; credit required; ship inside your app with conditions.
  Connect: download from https://huggingface.co/microsoft/VibeVoice-1.5B (no API).
  Rule: MIT-licensed weights, but Microsoft states the model is for research use and does not recommend commercial use.
  Rule: Every generated file carries an audible AI disclaimer and an imperceptible watermark.
  Rule: Not licensed for: voice cloning without recorded consent, disinformation, live deepfake voice conversion.
  Rule: Microsoft removed the TTS inference code from GitHub in September 2025; the 1.5B weights stay on Hugging Face.
  License: https://huggingface.co/microsoft/VibeVoice-1.5B (checked 2026-10-25; details: https://agentgid.com/assets/vibevoice/)
```

Not verified: The license text (MIT) allows commercial use, while the model card restricts intended use to research; commercial and in_app set to 'conditions' for that conflict.

## Links
- AgentGid page: https://agentgid.com/assets/vibevoice/
- Is it free for commercial use? https://agentgid.com/assets/vibevoice/commercial-use/
- Source: https://huggingface.co/microsoft/VibeVoice-1.5B
- Source: https://github.com/microsoft/VibeVoice
- Source: https://github.com/microsoft/VibeVoice/blob/main/LICENSE


## VibeVoice: research voice model for long podcast-style dialogue

VibeVoice is a family of free voice models from Microsoft Research. The 1.5B model turns a written script into spoken conversation up to 90 minutes long with up to 4 voices, in English and Chinese. Microsoft limits it to research and removed the ready-to-run speech code, so it is for experiments, not for a product.

### Where to use it
- Research on podcast-style dialogue: Study how a model keeps several voices apart over a long script, for example a two-host talk show written as turns with speaker labels.
- Long narration in one pass: Read a long text, such as a chapter or a lecture script, in one consistent voice instead of stitching many short clips together.
- Let a chatbot start speaking early: With the small Realtime model, a chatbot answer can be read aloud while it is still being written, in English and with one voice.
- Read a live data feed aloud: Turn a stream of short updates, such as scores or status messages, into speech for a research prototype.
- Compare voice models for a decision: Listen to VibeVoice next to models with product-friendly licenses before you pick one for your app. (idea, not from the official pages)
- Test text clean-up before speech: Check how your app should clean text, such as numbers, symbols and code, before any voice model reads it. (idea, not from the official pages)

### How to start
- Use a machine with an NVIDIA graphics card (Microsoft names the T4) or a fast Apple Silicon Mac. Microsoft recommends its NVIDIA container for the setup.

```
sudo docker run --privileged --net=host --ipc=host --ulimit memlock=-1:-1 --ulimit stack=-1:-1 --gpus all --rm -it  nvcr.io/nvidia/pytorch:24.07-py3
```

- Get the code from GitHub and install it with the streaming extra. Only the Realtime 0.5B model has working install steps; the long-form 1.5B speech code was removed.

```
git clone https://github.com/microsoft/VibeVoice.git
cd VibeVoice/

pip install -e .[streamingtts]
```

- Run the file example; the model downloads from Hugging Face on first run.

```
python demo/realtime_model_inference_from_file.py --model_path microsoft/VibeVoice-Realtime-0.5B --txt_path demo/text_examples/1p_vibevoice.txt --speaker_name Carter
```


### License
The code and model files are under the MIT license, so you keep the copyright notice with copies. But Microsoft says the model is for research only and does not recommend it in paid or real-world products.

### Good to know
- Every file it makes carries a spoken AI notice and a hidden watermark. You cannot get clean voice audio for a product this way.
- Copying a real person's voice without recorded consent, fake recordings and live voice changing are outside what the release allows.
- The long-form 1.5B speech model has no official way to run it now. Its install steps say they were turned off after misuse.
- It speaks only English and Chinese well, and only speech. It does not make music or sound effects on purpose, and it can add background music by chance.
- The small Realtime model has one voice at a time and English only. Voices come built in; for your own voice you must contact Microsoft.

Checked 2026-11-10 against: https://huggingface.co/microsoft/VibeVoice-1.5B, https://huggingface.co/microsoft/VibeVoice-Realtime-0.5B, https://github.com/microsoft/VibeVoice, https://github.com/microsoft/VibeVoice/blob/main/README.md, https://github.com/microsoft/VibeVoice/blob/main/docs/vibevoice-tts.md, https://github.com/microsoft/VibeVoice/blob/main/docs/vibevoice-realtime-0.5b.md, https://github.com/microsoft/VibeVoice/blob/main/LICENSE, https://github.com/microsoft/VibeVoice/blob/main/pyproject.toml, https://huggingface.co/api/models/microsoft/VibeVoice-1.5B, https://pypi.org/pypi/vibevoice/json

### Mistakes to avoid
- Don't: Looking for install steps or a script that runs the 1.5B long-form model from the official GitHub code → Do: Only the Realtime 0.5B model has official install and run steps; treat the 1.5B weights as research material (Microsoft removed the long-form speech code and marks its install section as disabled.)
- Don't: pip install -e . with the newest transformers for the Realtime model → Do: pip install -e .[streamingtts], which pins transformers 4.51.3 (The streaming extra fixes the transformers version the Realtime model was built against.)
- Don't: Feeding raw chatbot output with code blocks, formulas or symbols straight into the Realtime model → Do: Strip or rewrite code, math and unusual symbols to plain words first (The model does not read code, formulas or uncommon symbols and gives unpredictable audio.)
- Don't: Sending one- to three-word replies such as "OK, done." on their own → Do: Join very short replies with the next sentence before speaking them (Stability drops when the input is three words or fewer.)
- Don't: Using VibeVoice audio as the shipped voice of a product or a public video → Do: Use it for research, and pick a model with a product-friendly license for the shipped voice (Microsoft limits intended use to research, and every file carries an audible AI notice.)

### Questions
- Can I use VibeVoice for the voice in my paid app? Microsoft does not recommend it. The license is MIT, but the model card limits the model to research and says it is not for commercial or real-world use without more work.
- Can I still run the 90-minute model? The weights are still on Hugging Face, but Microsoft removed the speech code from GitHub and marks its install steps as disabled.
- How many voices can talk in one file? Up to 4 with the 1.5B model. The Realtime 0.5B model has one voice at a time.
- Which languages work? The 1.5B model is trained on English and Chinese only; the Realtime model is meant for English, with nine more languages marked as experimental.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
