VoiceStudio Passes 71,000 GitHub Stars as a Local ElevenLabs Rival, but Its Default Voice Model Is Non-Commercial

VoiceStudio, an open-source desktop app that calls itself "the open-source, fully-local ElevenLabs alternative", gained 2,301 stars in a day on GitHub trending. Anyone planning to sell what it produces has a licence question to settle first.

The repository, owned by a user named debpalash, shows 71.2k stars, 7.8k forks and 4,214 commits. The page lists 87 issues and 114 pull requests. As the Terminal wrote about REA's 270,097 stars, a star count measures attention, not users.

What the app does and what it runs on

VoiceStudio handles voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation. It is an Electron app with a local API and MCP integration, meaning an agent can call it as a tool. The README cites 646 languages.

Acceleration depends on the machine. NVIDIA GPUs use CUDA on Windows and Linux, and Apple Silicon uses Metal. With no dedicated GPU it runs on the CPU, slower, after installing a small CPU build of PyTorch that needs about 5 GB of free disk. Windows on ARM is marked experimental and CPU-only, and an Intel Mac can only run the interface against a remote backend.

Version 0.5.3 was the last release built on Tauri, the README says. Electron is now the only desktop app, so existing Tauri users must install it separately.

Two licences, and the second one binds

The application is AGPL-3.0. The README says that licence "covers the application", and then separates model weights, tokenizers, voices and generated-output conditions, which have their own terms. It states that a paid app licence does not grant model rights, and that the Model Manager's reviewed-file preview "does not enforce ordinary downloads or certify commercial permission."

The only engine the README names is its default, "powered by k2-fsa/OmniVoice". The OmniVoice model page splits its own licence in two:

Part	Licence	Stated reason
OmniVoice code	Apache 2.0	None given
Pre-trained model	CC-BY-NC	"constraints from its training data (e.g., Emilia)"

CC-BY-NC forbids commercial use. The README does not say which checkpoint the default engine downloads, so the Terminal cannot confirm the weights VoiceStudio fetches are the CC-BY-NC ones. The Model Manager, which this reporting could not open, is where the app says that is disclosed.

What OmniVoice claims

OmniVoice has 0.6B parameters and, per its card, over 600 languages; the page header lists 646. Its authors, who include Han Zhu and Daniel Povey, claim "state-of-the-art voice cloning quality from a short reference audio" and a real-time factor as low as 0.025, about 40 times faster than real time. Both are developer-supplied figures. The card showed about 1,576,534 downloads last month and 1.55k likes.

Open weights do not all come with the same freedom: Reflection's 501B Beam model shipped under Apache 2.0, while the OmniVoice card carries a disclaimer: "Users are strictly prohibited from using this model for unauthorized voice cloning, voice impersonation", and from fraud or scams.

Consent is a request, not a control

VoiceStudio's README says to "Clone voices only with permission." It describes no technical check for consent, such as verifying that the reference voice belongs to the person using it. Usage analytics need user consent, and the install guide for agents covers asking before model downloads.

The practical reading is narrow. A podcaster cloning their own voice for personal projects is inside what both pages describe. A studio selling dubbed video made with the default engine would need to read the OmniVoice licence first.

The next thing to check is the Model Manager's per-model notices, where the README says unresolved review gaps are listed. The repository's 87 open issues are where those gaps would surface.