VibeVoice: Summary of the Community License and Forks, The Future, and Downloading VibeVoice

[Arcitec]

Part 1: Confirmation that the VibeVoice-Community repo contains the final, official Microsoft VibeVoice code and the full Git history.

The most important aspect of continuing a community fork is to verify that the code is legitimate. I have now done that work.

Since GitHub keeps the Activity log even if they erase a repo, I had a look at the official Microsoft VibeVoice activity:

https://github.com/microsoft/VibeVoice/activity?ref=main

The thing that stood out the most is that 8 days ago they started spamming a bunch of "Update branch" without any informative titles. This was when they started nuking stuff.

Those commits have been garbage collected and are marked as "Missing commit", which on GitHub is a manual action by GitHub staff (otherwise nothing is ever truly deleted). So what they did was basically that they committed these "Update branch"-titled commits which deleted the files, and then they deleted the commits themselves, so that it wouldn't show the "lines of code that were being erased", thinking that it was enough to hide the original data. And finally they did a force-push with the new history which was the emptied repo with the statement.

First of all, that's a lot of very stupid effort for a repo that is already forked by over 1000 people.

Anyway, then you just have to keep scrolling down in the Activity log, and you suddenly see the last real commit, 9 days ago, one day before the nuking began: "update Colab: add download audio".

Then click the "..." on the right and go to "Compare changes". This brings you to a URL showing the full hash of the commit:

microsoft/VibeVoice@4b5b8be...6065c52

And lastly, the final magic: We can browse the repo contents that existed at that commit, since we now know the full hash:

https://github.com/microsoft/VibeVoice/tree/6065c5224e1fee5b2eaeef15f79ee8ef7de947d0

There is no way to clone such an "orphaned" commit, but there is a way to download the repo contents at an exact commit. So let's download it directly. Aaaand, voila - the final official MIT-licensed VibeVoice code:

https://github.com/microsoft/VibeVoice/archive/6065c5224e1fee5b2eaeef15f79ee8ef7de947d0.zip

---

Next, I had a look at this community fork:

git clone https://github.com/vibevoice-community/VibeVoice.git VibeVoice-Community

The first thing that I immediately saw is that git log reveals the 6065c5224e1fee5b2eaeef15f79ee8ef7de947d0 commit, which confirms that you continued your repo's history from the last official upstream commit.

Then I also ran a recursive diff to see which files differ:

diff -rq VibeVoice-6065c5224e1fee5b2eaeef15f79ee8ef7de947d0/ ../VibeVoice-Community/

Only in ../VibeVoice-Community/: EXAMPLES.md
Only in VibeVoice-6065c5224e1fee5b2eaeef15f79ee8ef7de947d0/Figures: Google_AI_Studio_2025-08-25T21_48_13.452Z.png
Files VibeVoice-6065c5224e1fee5b2eaeef15f79ee8ef7de947d0/README.md and ../VibeVoice-Community/README.md differ
Only in VibeVoice-6065c5224e1fee5b2eaeef15f79ee8ef7de947d0/: SECURITY.md

---

To sum it up:

1. Yes, VibeVoice-Community is a clean fork of the last official commits.
2. It then erases a useless screenshot, a SECURITY.md file which was just a Microsoft "please report security vulnerabilities to us" notice, and edits the README a bit to mark it as a community fork under the same MIT license as Microsoft's original code. The fork is completely legal and there's nothing that can be done legally to take it down, since Microsoft released all code and the models under the MIT license.
3. Since this repository is a clean fork with the exact same Git hashes, the entire VibeVoice development history is available here in the git log, which is great news.

---

Let's hope the community continues developing the project from here. It's up to us now!

There's already over 100 forks of this VibeVoice-Community repo. So everyone, do your part and click Fork and get involved here! ❤️

Here is a direct link to create a fork of VibeVoice-Community:

👉 https://github.com/vibevoice-community/VibeVoice/fork

Don't forget to ⭐ Star the VibeVoice-Community repository too! Get involved, spread the word and let's build up the community again ourselves!

There is also a new community discord to organize VibeVoice-Community development! Welcome!

👉 https://discord.gg/ZDEYTTRxWG

[Arcitec]

Part 2: Where to download the VibeVoice models.

Microsoft deleted the models from HuggingFace but they still exist on ModelScope.

Download the models from here:

• https://www.modelscope.cn/models/microsoft/VibeVoice-Large
• https://www.modelscope.cn/models/microsoft/VibeVoice-1.5B

Here's how to download the models:

1. Install the uv package manager. The simplest way for lazy people is pip install uv.
2. Run modelscope to download the models:

uvx modelscope download --model microsoft/VibeVoice-Large --local_dir ./modelscope/microsoft/VibeVoice-Large
uvx modelscope download --model microsoft/VibeVoice-1.5B --local_dir ./modelscope/microsoft/VibeVoice-1.5B

There's also a useless "VibeVoice-Tokenizer" repo, but it's always been empty (just a blank config file). It's here if anyone is curious:

uvx modelscope download --model microsoft/VibeVoice-Tokenizer --local_dir ./modelscope/microsoft/VibeVoice-Tokenizer

After downloading, your directories will all contain two hidden files named .mv and .msc. They are just metadata for the ModelScope downloader. You can delete those if you want.

---

Here's a bit of a bonus investigation:

• VibeVoice-Large and VibeVoice-7B are the exact same model. Exact same file hashes for all files.
• VibeVoice-Large is the official name for it. That repo also improved one of its metadata files to tag it properly as a "text to speech" model instead of a "others" category.
• Therefore, don't bother downloading any "-7B" repos. This "-Large" repo is all you need.

---

There's also a community mirror of the models here. I have verified the model weights are identical:

• https://huggingface.co/aoi-ot/VibeVoice-Large
• https://huggingface.co/aoi-ot/VibeVoice-1.5B

---

Here are the official hashes I've generated from Microsoft's ModelScope models:

VibeVoice-1.5B.sha256:

b0d6db52f45dd5179b42217c55112aa2ee66e7ad7b40b0950833c9f6780a0f21  config.json
30458d769bcf25aa4e8fd30bbde901f817e382a49f7c7da8c4380dd97b616876  configuration.json
c5f0a61ddeaeb028e3af540ba4dee7933ad30f9f30b6e1320dd9c875a2daa033  model-00001-of-00003.safetensors
81c3891f7b2493eb48a9eb6f5be0df48d4f1a4bfd952d84e21683ca6d0bf7969  model-00002-of-00003.safetensors
cb6e7e5e86b4a41fffbe1f3aaf445d0d50b5e21ed47574101b777f77d75fa196  model-00003-of-00003.safetensors
067db9b10fdecee3a5588aa00206794156c7125f5e85f3f2234e0e6d821ee629  model.safetensors.index.json
56bef7fb56db168e31dff685337c4acb6a35b5ce9c68d255128b01395b442c43  preprocessor_config.json
af44532437555b5fbfe9c28fbbd9e4d1a4793d6399d8055e7ee3db25aca7ec7d  README.md
64464f28380f81e76c88d76431a08b48c7f82a283e17f2e32c241c4b03407d19  figures/Fig1.png

VibeVoice-Large.sha256:

695598158e43b44227bc7aa6fd851e410f7ce30b21a5ea5c3fe22983961e500a  config.json
30458d769bcf25aa4e8fd30bbde901f817e382a49f7c7da8c4380dd97b616876  configuration.json
ae28d5c8f3587b518c7e371e96ebb69f74d854a854119acf433952bbc1926325  model-00001-of-00010.safetensors
c56b1ca707e31e435ded8b03baa4938d88275bf0ba7033935a16d8173a99ff85  model-00002-of-00010.safetensors
48bfb4af453d45e488050e90d3f39da0189f1c10a77d75223c2c2ced8b035baa  model-00003-of-00010.safetensors
b4893be477be68e53b8a9616422b99065f3d1431cce9efe0a1653495e9cf4df6  model-00004-of-00010.safetensors
471690e9846e791def400fefa3d2103c9839dc8a3e987b175f6539c7412422d6  model-00005-of-00010.safetensors
a7918d400ba895b15a1126fde242028e5d05b37bab0c0427944de81df80f901f  model-00006-of-00010.safetensors
b4f00ebea5a9f76eea891b3457621955433149ae603d921afa1498e46683ba37  model-00007-of-00010.safetensors
cc4b6fce97b76e847c742b59ab9463fd04b6d9fa69fc33e747ff722c2ab8cc28  model-00008-of-00010.safetensors
824db8970518950117f0d6ed859740d973b9436718c24f29bc78854c4587a4b2  model-00009-of-00010.safetensors
bc76bba7a46a0a748cc169efb6ccfb7617881e0fc3b533f67887d5957e1836e3  model-00010-of-00010.safetensors
dbcfc6e307494bc87684471872f3d8b785cb68b3589b6b306c43fde629b88ebd  model.safetensors.index.json
5a26081a18cd60f48d7ed36b904e68c24271ba9711d6328b53f7ad3eed446cce  preprocessor_config.json
b4e1717f0006df47e285a1eaa951f5b6f7ca6ca81654913a0f982328836dbf7e  README.md
64464f28380f81e76c88d76431a08b48c7f82a283e17f2e32c241c4b03407d19  figures/Fig1.png

Save these hashes to text files in the model directories, and verify the hashes with sha256sum -c VibeVoice-Large.sha256.

[Arcitec]

Part 3: ComfyUI

For ComfyUI users, the best way to run it is via one of these packages:

• https://github.com/Enemyx-net/VibeVoice-ComfyUI
• https://github.com/wildminder/ComfyUI-VibeVoice

Both of them have integrated the latest official code and improved the code, and added support for lots of optimizations such as SageAttention. They run more efficiently than Microsoft's official code did.

I am not sure which one has the best-written code. I hope they will consider joining forces and merging the two projects into one so that they can become stronger together and focus the community on one project. I have proposed a merger here:

wildminder/ComfyUI-VibeVoice#56

---

Update: I also found this very impressive node which is a unified TTS toolbox for VibeVoice, Higgs Audio and others inside ComfyUI:

• https://github.com/diodiogod/TTS-Audio-Suite

That author has really good organization and clean release notes (see the VibeVoice announcement for an example) so it is probably the one with the best code, judging by the author's clean work style. There's also a reddit thread by the author.

I also had a quick check and noticed that it automatically uses SageAttention if installed and available in your ComfyUI environment.

---

I haven't had time to evaluate any of the ComfyUI nodes. So feel free to let me know your results if you've compared the code quality and features of these three ways to run VibeVoice in ComfyUI. But from what I've seen, I think TTS-Audio-Suite seems by far the best.

[Arcitec]

Part 4: Licensing and Community Rights

Have fun everyone, and thank God that the MIT license exists:

https://www.tldrlegal.com/license/mit-license

It allows us to modify, distribute, sublicense, privately use, and commercially use all the VibeVoice code and weights. The only requirement is that we don't hold Microsoft liable for any damages. And we must include their copyright and the MIT license in our derivative works.

Furthermore, since Microsoft released it under a modified MIT license, we must follow their additional rules. They released it as "MIT + these restrictions". The restrictions are basically: Do not clone voices without consent. Do not use the model for malicious purposes.

---

After things have been released under a license, it cannot be revoked by anyone whatsoever. Only newer releases can have a different license. What's already out in the world is therefore forever under the MIT license. This is why it's so important to always make the first commit in a repository be the LICENSE, so that all commits after that are covered by the intended license. It is also why projects that change their license can still be forked and developed under the old license at the commit that happened before the license change.

Microsoft VibeVoice's first-ever commit was hash 0a67c9a7183b8e5b5773a45a2918556ebaac7e3b on August 25th, 2025. Which contained a single file: LICENSE. The MIT license. This means that every commit after that date - the entire VibeVoice history - is MIT licensed.

So this repository is forever legal. Enjoy!

---

Microsoft's MIT license modifications are as follows (text copied from their original repo). Some of the things below are meant to be license restrictions. Others are only clarifications that the model is unable to do certain things (such as real-time audio or other languages, since it wasn't trained for that). Microsoft wrote it very sloppily, so I'll include their full text.

Out-of-scope uses:

Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by MIT License. Use to generate any text transcript. Furthermore, this release is not intended or licensed for any of the following scenarios:

• Voice impersonation without explicit, recorded consent – cloning a real individual’s voice for satire, advertising, ransom, social‑engineering, or authentication bypass.
• Disinformation or impersonation – creating audio presented as genuine recordings of real people or events.
• Real‑time or low‑latency voice conversion – telephone or video‑conference “live deep‑fake” applications.
• Unsupported language – the model is trained only on English and Chinese data; outputs in other languages are unsupported and may be unintelligible or offensive.
• Generation of background ambience, Foley, or music – VibeVoice is speech‑only and will not produce coherent non‑speech audio.

I have emphasized the "this release is not intended or licensed" portion to highlight why this is released under a modified MIT license. When the copyright holder says that it is not licensed for certain uses, then their word is law, as long as they said it before/during the release of code and weights. That is what they did. So their extra restrictions are binding. We are not protected by the MIT license in those cases, because Microsoft explicitly listed forbidden use-cases where they do not give us the right to use it under the MIT license. In those forbidden cases, the code/weight license reverts back to the legal default instead - the owner's copyright protection - which means that Microsoft alone has the right in those cases.

We could argue about how sloppily written their extra restrictions are and how they would have been clearer if they had been included in the LICENSE file instead, but nonetheless, their intention was to limit some uses and they have the right to do that. Lawyers might want to argue that it's not binding because Microsoft didn't place the limitations inside the LICENSE file, which creates some ambiguity, but they would most likely prevail in court.

When companies use an open source license but place extra restrictions on it, it's known as a "fauxpen source license". Because the full, unrestricted open source license no longer applies to the project.

Thankfully, VibeVoice mostly added extra restrictions related to evil usage. But some things may need further analysis to ensure we stay out of infringing/unlicensed usage. The biggest issue is the sloppy instructions. They did not clearly separate the two lists of "missing features" vs "forbidden features", so they have mixed legitimate non-licensed usage together with general limitations due to features they were going to add later themselves, like a realtime streaming mode. So they do not intend to limit realtime streaming/low latency innovations, because they already publicly said that such a feature would be coming. They only wanted to limit real-time voice deepfaking (where someone actually speaks and converts their own voice in realtime via VibeVoice). They've also said that other languages would be coming, so "only English and Chinese" usage was not meant as a license restriction, but just as a "out of scope usage".

---

From their list of extra restrictions, the clear-cut unlicensed usage that we MUST stay away from is as follows:

1. Do not impersonate real people's voices without explicit, recorded consent.
2. Do not create disinformation audio that are being presented as genuine recordings. Clearly labeled/obvious parodies are fine.
3. Do not create a real-time/low-latency voice-2-voice model from it, which would allow "live deep-fake" applications where someone speaks and another voice comes out to the listener. For example, we can't use a speech-to-text model to generate text that we feed into VibeVoice's text-to-speed, if we achieve real-time speeds. It is allowed if we achieve it with high latency, since the license only restricts low-latency results. They did not define how much latency is high/low, though, so a conservative estimate should be used to stay out of legal trouble if anyone wants to create a voice-2-voice system from this model.
4. General reductions of VibeVoice latency (even realtime) are permitted as long as it isn't being built as part of a voice-2-voice processing chain.
5. We can ignore the two notices about "unsupported languages" and "background music", because those points were only there to say that it may produce gibberish/non-coherent audio in those cases.

In short: The VibeVoice license does not meet the OSI’s Open Source Definition. So while Microsoft calls it "MIT", it’s really a non-standard license variant, which has implications for distribution and adoption. Thankfully, the restrictions are very minor and mostly target malicious usage. They unfortunately muddled restrictions and clarifications together, which creates ambiguity. But it is very clear that their intention was only to limit harmful misuse/abuse (deepfake disinformation, impersonation, live voice-swapping).

---

TL;DR:

• VibeVoice is licensed under a custom, non-OSI-compliant "fauxpen source" license.
• The safest legal interpretation is:
  • Treat non-consensual impersonation, disinformation, and low-latency voice-conversion as forbidden uses.
  • Treat language support and non-speech generation as capability limitations, not restrictions.
  • Treat latency reductions (even realtime) as permitted, as long as you aren't building a voice-2-voice processing chain.
• Legally, because these additional license restrictions were not placed in the LICENSE file, there is room for debate. However, any reasonable court would likely side with Microsoft, since the restrictions were present from the first commit and clearly reflect intent - and intent matters in court.

This is not legal advice; consult a lawyer for formal guidance if you intend to challenge Microsoft's license.

[Arcitec]

TL;DR Summary:

• This fork and all other forks and derivative works are legal forever, as long as they also follow the MIT license and Microsoft's extra restrictions ("don't be evil", don't clone without consent, don't build tools for malicious purposes).
• The last released, official weights are also legal forever. They are under the same license.

[Arcitec]

My hope is that the community rallies around this fork, since we're on our own. Time to put our thinking hats on and carry on the development ourselves!

Here is a direct link to create a fork of VibeVoice-Community:

• https://github.com/vibevoice-community/VibeVoice/fork

[chibiace]

2. The fork is completely legal and there's nothing that can be done legally to take it down,

microsoft owns this platform, might be worth making a mirror on another platform and keeping some backups offline.

just in case. :)

[Arcitec]

Definitely. Everyone should be git cloning this repository to make backups.

But I doubt that they'd be stupid enough to do the PR disaster move of revoking a completely legal repo which follows the exact license they released the original code under.

That would bring a lot of funny headlines in the open source world, and could get Microsoft into legal trouble with the Free Software Foundation and others.

[fakerybakery]

2. The fork is completely legal and there's nothing that can be done legally to take it down,

microsoft owns this platform, might be worth making a mirror on another platform and keeping some backups offline.

just in case. :)

Great idea. I have a backup locally myself. If it is removed I will move it to either a self-hosted git instance or codeberg/gitlab

[Arcitec]

@fakerybakery That is great to hear. I'll be with you on moving to GitLab if necessary. This is MIT code so it's untouchable by Microsoft.

By the way, Microsoft officially released these models as -1.5B and -Large.

They clearly considered labeling the large one as -7B first (since they released the initial repo with that name before renaming it to Large). But I think they then decided internally that they are not going to open source any more/larger models, so they renamed it to -Large. That was the final name before they stopped publishing the VibeVoice project.

I personally like the -7B because it says the exact parameter count. And I see that you adopted that as your preferred name too. But now it causes confusion because there's two repos from Microsoft. They are basically identical, except a tiny difference in the metadata file (the -Large file is newer and has more filled-in metadata, which has no effect on generation).

All this being said, two other major community projects already settled on keeping the Microsoft -Large and -1.5B names:

• https://github.com/Enemyx-net/VibeVoice-ComfyUI
• https://github.com/wildminder/ComfyUI-VibeVoice

And community quantizations use names like VibeVoice-Large-Quant-4Bit now.

I think it would be good for consistency to rename the links in your README to point at the mirrors of the -Large repo, so that we are all on the same consistent idea and same files, so that all projects easily refer to the same model directories without confusion. And let's thank Microsoft for creating this annoying naming split in the first place. 😉

[fakerybakery]

FYI started an unofficial discord server in case people want to discuss the community repo:
https://discord.gg/ZDEYTTRxWG

[Arcitec]

@fakerybakery Oh that is fantastic news. I'll update the Reddit thread with that link too.

https://www.reddit.com/r/StableDiffusion/comments/1ngfa9k/vibevoice_summary_of_the_community_license_and/

[ibrahimahtsham]

uhhh u are insanely legendary @Arcitec