VoiceStudio is a free, AGPL-3.0 desktop app with more than 14,450 GitHub stars that clones voices, dubs video, dictates, and transcribes across 646 languages on your own hardware, with no ElevenLabs credit meter, though its default voice model carries a noncommercial license worth checking first.
A marketing team dubbing product videos into several languages every month can watch its ElevenLabs bill move fast. On the ElevenLabs pricing page, automatic dubbing draws down credits at 2,000 to 3,000 per minute depending on whether the output keeps ElevenLabs' watermark, and Dubbing Studio, the mode that preserves individual speakers and lets someone fix timing by hand, runs 5,000 to 10,000 credits per minute. A single ten-minute video dubbed in Studio mode without a watermark can use nearly all of the 121,000 monthly credits on the $22-a-month Creator plan. VoiceStudio, an open-source desktop app that gained more than 800 GitHub stars in a day and has crossed 14,450 total, runs the same kind of dubbing pipeline, plus voice cloning, dictation, and transcription, on your own hardware with no credit meter at all.
What VoiceStudio actually does
VoiceStudio, previously named OmniVoice Studio, describes itself on GitHub as "the open-source, fully-local ElevenLabs alternative." That is not a marketing gloss added by a third party; it is the repository's own framing, and the project carries the elevenlabs-alternative topic tag directly.
The feature set matches the claim. Voice cloning is zero-shot from a short reference clip. Voice design builds a synthetic voice from attributes like age, accent, and pitch instead of a sample. The dubbing pipeline transcribes a video, translates it, keeps each speaker distinct, resynthesizes the dialogue, and exports the result. A dictation widget transcribes speech into any text field on the machine through a global hotkey. Under the hood, it routes work through 16 text-to-speech engines and 11 speech-to-text engines, and it advertises a 646-language catalogue, though the GitHub repository is upfront that actual language coverage and quality depend on which engine you pick.
It ships as a desktop app, built on a Rust shell, for macOS, Windows, Linux, and Docker, plus a local REST, SSE, and WebSocket API that is OpenAI-audio-compatible, and an MCP server for wiring the tool into coding agents or other software. None of that requires an account, an API key, or a subscription for the local workflow.
The application itself is licensed AGPL-3.0. Downloaded models are a separate matter: the repository's engine table notes that the default OmniVoice engine's pretrained weights are CC-BY-NC, meaning noncommercial use only, even though the surrounding application code is AGPL-3.0 and the model's own code is Apache-2.0. A team that wants to generate audio for commercial use with a permissively licensed engine has options in the same install, such as CosyVoice or GPT-SoVITS, but the default choice is the one with the license catch.
Why this matters for a business
The appeal for a business is the same one that shows up whenever a local tool displaces a metered SaaS product: nothing is billed per minute, per character, or per seat. A company that regularly localizes onboarding videos, product demos, or internal training material into multiple languages can move that entire workflow in-house and pay for hardware and someone's time to run it, rather than watching a single video eat a month's worth of dubbing credits.
The VoiceStudio product site also points at a real business gap the project hasn't closed yet: it lists "Commercial rights and hosted access" as a separate offering called VoiceStudio Pro, which is enquiry-only, with a Cloud version still in early access. The free, self-hosted tool is genuinely free. The path to a fully commercial, fully supported version of it is not yet a finished product.
Honest caveats
The GitHub page carries an active "Warning" banner: this is active beta software, the main branch may change between releases, and users are told to run the latest tagged release rather than main for stable work. Setup has real friction, too. The app needs 8 GB of RAM minimum, 16 GB recommended, and 10 to 20 GB of disk space for models. A GPU is optional but helps; the minimum useful VRAM is 4 GB, and the larger engines want 8 GB or more. Intel Macs cannot run the local backend at all and need a remote server instead. First launch downloads a default model, which takes real time depending on your connection. And because the app is AGPL-3.0 rather than a permissive license, a company that modifies VoiceStudio and offers that modified version to others as a network service is obligated to publish its changes under the same license, a constraint that a more casual MIT-licensed tool would not carry.
A quieter reading
"Free" and "no per-minute meter" are both true statements about VoiceStudio the application. They are not automatically true about every voice you generate with it. The more interesting story here is not that a free dubbing tool exists; it's that the license question in open-source voice AI has moved from the product's price tag down into the fine print of whichever engine you happen to select.