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
1. Price: huggingface.co/microsoft/VibeVoice-1.5B
Open weights, free to download.
huggingface.co/microsoft/VibeVoice-1.5BNo quote yet: price is read from the official pages.
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-1.5B3. Credit the author: LICENSE on GitHub
“The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.”
Speech research made with Microsoft VibeVoice (MIT license)
LICENSE on GitHub4. Agent access: github.com/microsoft/VibeVoice
There is no package or API. Files are downloaded by hand from the official site.
github.com/microsoft/VibeVoiceLicense 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
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
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
Mistakes to avoid
Microsoft removed the long-form speech code and marks its install section as disabled.
pip install -e . with the newest transformers for the Realtime modelpip install -e .[streamingtts], which pins transformers 4.51.3The streaming extra fixes the transformers version the Realtime model was built against.
The model does not read code, formulas or uncommon symbols and gives unpredictable audio.
Stability drops when the input is three words or fewer.
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.
“The VibeVoice model is limited to research purpose use exploring highly realistic audio dialogue generation” huggingface.co/microsoft/VibeVoice-1.5B
“The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.” LICENSE on GitHub
“We do not recommend using VibeVoice in commercial or real-world applications without further testing and development.” huggingface.co/microsoft/VibeVoice-1.5B
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
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.
- 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/)
Speech research made with Microsoft VibeVoice (MIT license)Alternatives
VibeVoice This page
Pick Chatterbox when you need an open voice model with an MIT license for a product, voice cloning and 23 languages.
Pick Kokoro for a small open model under Apache-2.0 with fixed voices that you can ship in an app.
Pick VoiceStudio when you want a desktop app for voices, dubbing and transcription instead of running model code.
Zero-shot voice-cloning TTS; code MIT, pre-trained weights CC BY-NC.
Family of small VITS text-to-speech checkpoints, one per language, covering over 1,100 languages.
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.