Quick answer: VoiceStudio is a free desktop app that runs voice cloning, video dubbing, dictation, transcription, and audiobook generation on your own hardware, with no account or API key. Download the latest release for macOS (Apple Silicon), Windows, or Linux; first launch sets up a managed Python environment and downloads the default model on its own.

If you keep seeing the name and are not sure what it refers to, one fact clears most of the confusion: this is the project previously called OmniVoice-Studio, renamed. Coverage of OmniVoice from earlier in 2026 describes the same app, and the rename is still visible in the plumbing — the official Docker image is palashdeb/omnivoice-studio, and the bundled Python package is still named omnivoice.

What it actually does

Six workflows: voice cloning and voice design, video dubbing, dictation, stories, audiobooks, and batch generation. Under them sit 16 text-to-speech engines and 11 speech-recognition engines you can switch between from a model catalogue. The headline “646 languages” is the size of that catalogue across engines, and the README itself adds the caveat that actual coverage and quality depend on the engine you select. Cloning a voice takes a sample of about three seconds, with 5 to 15 seconds recommended.

It is not only a GUI. The backend exposes a local REST/SSE/WebSocket API, an OpenAI-compatible audio API, and an MCP server, so agents and existing OpenAI-client code can drive it.

What it runs on

macOS 13.3+ on Apple Silicon only — Intel Macs cannot run the local backend and need a remote one. Windows 10/11 x64 via MSI, Linux x86_64 with glibc 2.39+, and Docker profiles for CUDA, ROCm, and CPU. Compute can be CUDA, Apple MPS/MLX, ROCm, plain CPU, or optional remote workers. The README carries an active-beta warning and points stable work at tagged releases.

The license, precisely

The part most coverage skips. The application is AGPL-3.0, which permits commercial use of the app itself, and the project sells a separate commercial license for embedding it in closed-source products. The default model is a different story: k2-fsa/OmniVoice’s model card lists its code as Apache-2.0 but its pretrained weights as CC-BY-NC, and the audio tokenizer carries separate Boson Higgs Audio 2 and Meta Llama community terms. “Free and open-source” is accurate for the software; whether your use of the default voices is commercial is a question those weight terms decide, not the AGPL.

Against ElevenLabs

The README’s comparison table is more careful than the “ElevenLabs alternative” framing around the project: it compares against “a typical hosted voice service” and claims no quality parity. The trade it describes is managed cloud compute versus local control — hosted services win on setup and scaling, VoiceStudio wins on data staying local, zero marginal cost after hardware, offline use, and swappable engines. We put the two side by side, feature by feature and number by number, in VoiceStudio vs ElevenLabs.

Repo, measured on September 6, 2026: created April 9, 2026, so the project is five months old; 18,820 stars with 5,150 gained in the preceding week per GitHub Trending; 2,108 files, 120 Markdown pages under docs/; latest release v0.5.1 (August 28, 2026); a Chinese README maintained alongside the English one.

Related: the agent-skills wave, measured covers how we read trending repos; transcribing and searching lectures is the neighboring workflow on the text side.

Disclosure: I build StashBase, a local searchable workspace; the two projects are unrelated. Everything above comes from the VoiceStudio repository, docs, and license files at a September 6, 2026 clone, plus GitHub API metadata from the same day. We have not yet run the app; nothing here is a usage report.