VoiceStudio does voice cloning, video dubbing, dictation, transcription and audiobook production, all on hardware you already own. No account, no API key, no usage meter. Tauri v2 shell, React interface, FastAPI backend, past 30,000 stars since April 2026.
It is good. The part most coverage skips is the part that can actually bite you.
Two licences, and the permissive one is not the one that matters
The application is AGPL-3.0. Run it, modify it, use it internally, all fine. Offer a modified version as a network service and you owe your source under the same licence.
The default engine is the one to watch. OmniVoice ships as Apache-2.0 code with CC-BY-NC weights. Non-commercial. So the out-of-the-box configuration, the one you get if you install it and start generating, produces audio from non-commercially-licensed weights.
If you are making anything commercial, switch engines deliberately. Several are Apache-2.0 end to end:
| Engine | Languages | Clones | Licence |
|---|---|---|---|
| OmniVoice (default) | 600+ | Yes | Apache-2.0 code, CC-BY-NC weights |
| CosyVoice 3 | 9 + 18 dialects | Yes | Apache-2.0 |
| VoxCPM2 | 30 | Yes | Apache-2.0 |
| MOSS-TTS-v1.5 | 31 | Yes | Apache-2.0 |
The project documents this. The write-ups are what gloss over it.
646 languages is a catalogue size, not a promise
The headline number comes from the default engine covering 600 or more. The other fifteen engines are nowhere near it: CosyVoice 3 handles 9 plus 18 dialects, VoxCPM2 covers 30, dots.tts 24, Supertonic 31.
So "646 languages" really means one engine has a very large catalogue. Pick the engine for your language and check that one specifically.
Cloning is zero-shot, which saves you effort
The reference clip is a prompt, not training data. Nothing is fine-tuned on your voice. The model listens to a short sample and conditions on the timbre it hears, which is why it finishes in seconds.
- Three seconds of audio is enough to work
- Five to fifteen seconds sounds noticeably better
- A cleaner clip beats a longer one, every time
If you were planning to record twenty minutes of reference audio, you do not need to.
The bit developers will care about
The backend exposes an OpenAI-compatible audio API. Migrating is one line:
- base_url="https://api.openai.com/v1"
+ base_url="http://localhost:3900/v1"
Anything written against OpenAI's speech endpoints keeps working, now against local compute at no per-request cost.
One ask: it embeds AudioSeal watermarking by default. Leaving it on costs you nothing, so leave it on, and clone a voice only with the consent of the person it belongs to.
Full write-up with all sixteen engines, the dubbing pipeline, hardware requirements and the hosted comparison: VoiceStudio: local voice cloning and dubbing in 646 languages.
For audiobook work I run scripts through a word counter to estimate runtime first, and an AI proofreader to catch typos that TTS will otherwise read aloud with total confidence.
Top comments (0)