VoiceStudio Has 75,300 Stars and an AGPL Licence, but Its OmniVoice Model Is CC-BY-NC

VoiceStudio, an AGPL-3.0 desktop app for voice cloning and dubbing, has 75,300 GitHub stars, while the speech model under its default engine is licensed CC-BY-NC, a non-commercial licence. The app's README says a paid app licence "does not grant those rights."

The repository belongs to a developer who goes by debpalash. Its README describes "Open-source voice cloning, voice design, video dubbing, dictation, transcription & audiobook creation in 646 languages," and the repository page calls it a "fully-local ElevenLabs alternative." The same page lists 8,200 forks, installers for Windows, macOS and Linux, and a Docker image.

The licence stack is the story

The README says "AGPL-3.0 covers the application, including commercial use under its conditions." It then says "Model weights, tokenizers, voices and generated-output conditions are separate." The Model Manager, it adds, "does not enforce ordinary downloads or certify commercial permission."

That is the passage to read twice. The app's default engine is described as "powered by k2-fsa/OmniVoice," and OmniVoice's own model card on Hugging Face lists the code as Apache 2.0 and the pre-trained model as CC-BY-NC. The card attributes the non-commercial term to constraints from its training data and names Emilia, a speech dataset, as an example.

Layer	Licence as stated	Source
VoiceStudio app	AGPL-3.0	VoiceStudio README
OmniVoice code	Apache 2.0	Hugging Face model card
OmniVoice pre-trained model	CC-BY-NC	Hugging Face model card

I could not confirm from the README which weights the installer downloads by default, so I cannot say that a given install is non-commercial. What the README does establish is that the app's licence and the weights' licence are separate documents, and that the maintainer says so. Anyone cloning a voice for a paid product has to read the model card, not the repository badge.

A model with 600 languages and an app with 646

The OmniVoice paper, arXiv:2604.00688, was submitted on April 1, 2026 and revised on April 21. Its authors, including Han Zhu and Daniel Povey, write that the model "scales to over 600 languages" and was trained on 581,000 hours of multilingual data they describe as curated entirely from open-source sources. The model card lists 0.6B parameters and 1,730,352 downloads in the last month.

The figure 646 appears in the VoiceStudio README and in the model card's tag list, while the paper and the card's prose say "over 600." Neither the README nor the paper says how well any of those languages perform. A language count is a count of tags until someone publishes per-language error rates.

1. VoiceStudio (app)75K stars
2. OmniVoice (k2-fsa)15K stars

Source: GitHub pages for debpalash/VoiceStudio and k2-fsa/OmniVoice, accessed 2026-10-19

Stars are attention, not adoption, as with CLI-Anything at 54,400 stars. The app has five times the stars of the research repository it depends on, and the research repository is the one whose model has 1.7 million monthly downloads.

A rename, a CLA and a Pro edition nobody has announced

Release titles through v0.4.2 say "OmniVoice Studio." From v0.5.0, dated August 14, they say "VoiceStudio." The old name survives in configuration: the v0.5.6 notes, dated September 23, still tell users to move the backend with the OMNIVOICE_PORT setting.

On October 2 the owner opened issue #2586, titled "Contributor License Agreement: sign here." It has 44 comments. The issue says the agreement covers "commercial/proprietary licensing" and that contributors "keep your copyright." A contributor, psiberfunk, read CLA 1.0 as granting "perpetual, irrevocable rights for proprietary products and services," and proposed that contributors receive a personal licence to any commercial or Pro edition containing their work. That is one contributor's reading, and the page I read does not show the proposal accepted or a Pro edition announced.

Licence terms that sit apart from the code are not new: Meta's Muse Spark open weights carry their own resale threshold. An AGPL project that asks contributors for relicensing rights is usually keeping the option of selling a non-AGPL version. The README's phrase "a paid app licence" points the same way. Neither document says what is being sold.

What installing it looks like on October 18

The v0.5.6 notes add the ability to answer Twilio phone calls in a saved voice, off by default. The v0.5.3 notes, from September 17, say the Tauri shell is retired and Electron is the only desktop app.

The newest open issues, all from October 18, are mostly failed installs: a backend that "did not answer on port 3900 within 300 s," a missing Python environment under C:\Windows\System32\VoiceStudio, and a model that fails to load on CachyOS. Several users have posted the same "[Install & FIX]" guide. The README offers a one-line curl ... | sh installer for macOS and Linux and irm ... | iex for Windows, which pipe a remote script straight into a shell. I did not run either.

If you want to dub your own videos on a machine with an NVIDIA GPU or Apple Silicon, the AGPL app is the easy part. If the output goes into something you sell, the sticking point is the weights, and the README tells you to go and check them. The next dates to watch are the CLA 1.0 text and whether a Pro edition appears in the release notes after v0.5.6.