Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026

VibeVoice

by Microsoft Research

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

Voiceover (text-to-speech) Open source

GitHub repoWebsite

2. Commercial use: huggingface.co/microsoft/VibeVoice-1.5B

“The VibeVoice model is limited to research purpose use exploring highly realistic audio dialogue generation”

The MIT license allows it, but Microsoft limits intended use to research and does not recommend commercial use.

huggingface.co/microsoft/VibeVoice-1.5B

License checked Oct 25, 2026 against 2 official pages. Tap a verdict to see its source. How we checked Report an error

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 you can 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 Idea

Listen to VibeVoice next to models with product-friendly licenses before you pick one for your app.

Test text clean-up before speech Idea

Check how your app should clean text, such as numbers, symbols and code, before any voice model reads it.

As reported by Microsoft Research: Up to 90 minutes of speech with up to 4 distinct speakers; The small Realtime model gives first audible speech in about 300 ms (hardware dependent).

VibeVoice popularity

52 / 100
Gid Score, same scale as developer tools
57,295
GitHub stars
VibeVoice: GitHub starsmicrosoft/VibeVoice 57,295

How the Gid Score is calculated

How to start

Tell your AI assistant

“For a research test only, install microsoft/VibeVoice from GitHub with the streamingtts extra and read my text file with VibeVoice-Realtime-0.5B; do not use it in the product.”

Show codeStep 1 or Step 2 or Step 3

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

Mistakes to avoid

Looking for install steps or a script that runs the 1.5B long-form model from the official GitHub code
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.

pip install -e . with the newest transformers for the Realtime model
pip install -e .[streamingtts], which pins transformers 4.51.3

The streaming extra fixes the transformers version the Realtime model was built against.

Feeding raw chatbot output with code blocks, formulas or symbols straight into the Realtime model
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.

Sending one- to three-word replies such as "OK, done." on their own
Join very short replies with the next sentence before speaking them

Stability drops when the input is three words or fewer.

Using VibeVoice audio as the shipped voice of a product or a public video
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.

License

MIT (model card limits intended use to research) 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.

Commercial use
With conditions

“The VibeVoice model is limited to research purpose use exploring highly realistic audio dialogue generation” huggingface.co/microsoft/VibeVoice-1.5B

Credit
Required

“The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.” LICENSE on GitHub

Inside your app or game
With conditions

“We do not recommend using VibeVoice in commercial or real-world applications without further testing and development.” huggingface.co/microsoft/VibeVoice-1.5B

Pass the files on to your users
With conditions

MIT lets you share copies if the copyright and permission notice go with them.

“to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software” LICENSE on GitHub

AI training
Not stated
Bots and bulk downloads
Not stated

AllowedWith conditionsNot allowedNot stated or not checked yet

Is VibeVoice free for commercial use? The full answer Not legal advice. Check the license page before you ship.

For your agent

License rules tell your agent what it may do with VibeVoice. Paste them into your prompt or AGENTS.md.

License rules
- 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/)
Can an agent use it by itself?
No.
API, key or account
Not needed. Everything runs on your side.
MCP server
None found
Credit line, if you want one
Speech research made with Microsoft VibeVoice (MIT license)
This card for machines

Alternatives

VibeVoice This page

Free With conditions Download

Chatterbox

Pick Chatterbox when you need an open voice model with an MIT license for a product, voice cloning and 23 languages.

Free Commercial OK PyPI

Kokoro-82M

Pick Kokoro for a small open model under Apache-2.0 with fixed voices that you can ship in an app.

Free Commercial OK npm, PyPI

VoiceStudio

Pick VoiceStudio when you want a desktop app for voices, dubbing and transcription instead of running model code.

Free With conditions MCP, API

F5-TTS

Zero-shot voice-cloning TTS; code MIT, pre-trained weights CC BY-NC.

Free With conditions PyPI

Meta MMS-TTS

Family of small VITS text-to-speech checkpoints, one per language, covering over 1,100 languages.

Free Non-commercial Download

All voiceover (text-to-speech)

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.