VoiceStudio Is AGPL, but Its Default Voice Model Is Non-Commercial

VoiceStudio, a voice-cloning and dubbing desktop app written by Palash Debnath, had 50,471 GitHub stars and 5,604 forks as of Thursday, and it is licensed AGPL-3.0, which allows commercial use. The default speech model it downloads is not: the OmniVoice weights carry a CC-BY-NC licence, which does not.

The project's license notice is direct about the split. VoiceStudio's own code is free for "commercial and internal business use", it says, but "model weights, tokenizers, and other third-party assets retain their own terms". The repository was created April 9, 2026, and its latest release is v0.5.6, published Sept. 23.

What the default engine is under the hood

The README says the default engine is "powered by k2-fsa/OmniVoice", a text-to-speech model from the Next-gen Kaldi group. Its Hugging Face card is the source for the licence terms: "Our code is released under the Apache 2.0 License. The pre-trained model is licensed under the CC-BY-NC due to constraints from its training data (e.g., Emilia)." The weights had 1,431,614 downloads when the page's API was read.

VoiceStudio's notice adds a second layer. The OmniVoice audio tokenizer ships with separate Boson Higgs Audio 2 and Meta Llama community terms, and the notice says a commercial licence for VoiceStudio's code "does not replace any of those terms".

Layer	Licence	Commercial use
VoiceStudio app code	AGPL-3.0	Allowed, with source disclosure for network use
OmniVoice code	Apache 2.0	Allowed
OmniVoice weights	CC-BY-NC	Not allowed
OmniVoice audio tokenizer	Boson Higgs Audio 2 and Meta Llama terms	Governed by those terms

A company that installs the app, clones a voice with the default engine and ships the audio in a product would be using non-commercial weights. The AGPL does not change that. The notice also offers a separate commercial licence for embedding VoiceStudio in closed-source products, and lists no price for it yet.

What it asks of the user

The README tells users to "clone voices only with permission", and says remote services are optional and that "usage analytics requires consent". The audio-quality documentation refers to a "provenance watermark setting" that stays in effect when the app's volume-levelling step is switched off. The documentation does not say whether the watermark is on by default.

Installation is one line, curl -fsSL https://voicestudio.sh/install | sh, which runs a script from the project's own domain. Binary downloads exist for macOS, Windows, Linux and Docker.

What the project's own numbers say

The performance guide gives timings from an Apple Silicon M2 with 16 GB of memory, and says hardware will differ but the ratios hold. Encoding a reference clip takes about 0.4 seconds and is cached afterward, and the first model load takes about 8 seconds. CPU-only runs are about 2x slower than on Apple's GPU path.

On NVIDIA cards with 16 GB or more of memory it allows up to three or four concurrent generations. Apple Silicon runs one at a time by design. These are the maintainer's measurements, not independent ones.

The issue tracker shows the cost of the growth. When read, it listed 77 open issues, among them #2445, "Backend did not answer on port 3900 within 300 s", and #2462, "Generation failed. Check the selected engine and try again."

OmniVoice's own README claims support for "600+ Languages", and VoiceStudio's headline says 646. The project has not published a release date for its commercial tiers. For the commercial side of voice A.I., see the report on ElevenLabs' $22 billion employee tender, and for another open-weight licence read closely, Qwen3.8-27B under Apache 2.0.