What is it?
VoiceStudio is a desktop app that runs voice models on your own machine: clone a voice from a reference recording, design one from a description, dub a video, produce an audiobook, dictate or transcribe. The repository describes it as an open-source, fully local alternative to ElevenLabs. The default engine, k2-fsa's OmniVoice, covers 646 languages according to the README.
For transcription alone, the comparison of local speech-to-text tools looks at dedicated tools, including Murmure for dictation and Scriberr for recorded files.
Why is it interesting?
- A catalog of engines: the docs list 18 speech engines (OmniVoice, CosyVoice 3, IndexTTS 2.5, VoxCPM2, gpt-sovits and more) and 10 transcription engines (WhisperX by default, Faster-Whisper, MLX Whisper, Parakeet TDT, Moonshine and more). Each one installs from the model manager.
- From dubbing to batch queues: video dubbing with timed speech, vocal isolation, speaker diarization, stories and audiobooks, batch jobs and watch folders.
- A local API and an MCP server: an agent or a script can drive generation without going through the interface.
- An invisible watermark: when AudioSeal is available, generated audio gets an inaudible mark that flags it as synthetic.
- Acceleration that fits the machine: CUDA on NVIDIA cards, Metal on Apple Silicon, opt-in ROCm for some Radeon cards on Windows, Vulkan for a few GGUF engines. Without a GPU everything runs on the CPU, with a CPU build of PyTorch of about 5 GB.
Good to know
The default model is non-commercial. The app is under the AGPL-3.0, commercial use included, but its licence notice states that the weights of OmniVoice, the default engine, are CC-BY-NC, and that its tokenizer falls under Higgs Audio 2 and Llama terms. A paid VoiceStudio licence grants no rights over those models. The model manager shows each model's licence, and the project does not yet claim commercial clearance for any of them.
Not every machine gets the same app. On an Intel Mac, the app is the interface only and must connect to a remote backend. Windows on ARM is experimental and CPU-only. On Windows, PyTorch GPU acceleration is NVIDIA-only, and the performance guide warns that a silent fallback to the CPU makes everything several times slower.
A Pro plan and a contributor agreement. Pro, at $99 per user per year or $299 for life, adds several devices per seat, remote devices and early access. A commercial licence covers embedding VoiceStudio in a closed product, and contributors sign a CLA in favour of Yupcha Softwares, the company that maintains it. Usage analytics, sent to PostHog, wait for the user's consent.
A project at version 0.5. The repository was created in April 2026. The latest stable release is 0.5.6 from 23 September, and 0.5.7 has been in beta since 23 October. The desktop app moved from Tauri to Electron after 0.5.3, so Tauri installs need a separate Electron install. The creator authored 3,980 of the 5,481 commits, and 199 issues and pull requests are open.