VoiceStudio

What it is

VoiceStudio is a free, open-source voice studio that runs on your own computer, in Docker or on a phone. Its README lists six jobs. It clones a voice from a clean reference recording and designs one from a written description. It dubs video into other languages, dictates into any text field, transcribes audio and narrates stories and audiobooks. The default engine is OmniVoice from k2-fsa, and a Model Manager offers dozens more. Palash Debnath created it, and a notice file says Yupcha Softwares Private Limited maintains it. The README says no paid cloud sits behind it, and the app asks before it downloads any model.

1. Install the app for your system
2. First launch builds its own Python environment
3. You pick a voice model and download it
4. Clone, design, dub, dictate or transcribe
5. Audio is saved in the app’s folders on your machine

Who it suits

People who make narration, dubs or transcripts and want the audio to stay on their own machine, plus developers who want a local speech server for an agent.

Skip it if you want a hosted service with an account and nothing to install. The default voice model’s weights are non-commercial too, so check the terms before you sell audio made with it.

Where to use it

The README’s platform table covers desktops, servers and phones, and the maker links an install guide for most rows.

• Windows 10 and 11 on x64, with NVIDIA CUDA or the CPU. Windows on ARM is marked experimental
• macOS on Apple Silicon, with Metal. Intel Macs get the app UI only
• Linux on x86_64: AppImage, .deb, Snap Store, an AUR package and a community Nix flake
• Docker: a headless backend with a web page at localhost:3900
• Android 10 and iOS 17 and later: preview apps that run models on the phone
• Agents and automations: an MCP server and an OpenAI-compatible speech API

How to set it up

Get it saves a zip of Mindket’s setup notes in your downloads folder, with one file per platform from the maker’s install guides.

1. Run the install line for your system. On Windows, use a regular PowerShell, not an administrator one.
2. Open VoiceStudio and click Install local runtime. It checks free disk first.
3. Go through the four first-run steps: System check, Model packs, Privacy and Enter studio. Models download only when you click.
4. Open Voice cloning, pick a voice or add a clean reference recording, type your text and generate.

curl -fsSL https://voicestudio.sh/install | sh

irm https://voicestudio.sh/install.ps1 | iex

If macOS says the app is damaged, the download is intact. The guide says release builds are not notarized, and this line clears the quarantine flag.

xattr -dr com.apple.quarantine "/Applications/VoiceStudio.app"

Docker serves the web page instead of the desktop app. Make an API key first. The guide says the first run downloads about 2.4 GB of weights.

export OMNIVOICE_API_KEY="$(python3 -c 'import secrets; print(secrets.token_urlsafe(32))')"

docker run -d --name omnivoice \
  -p 127.0.0.1:3900:3900 \
  -e OMNIVOICE_API_KEY="$OMNIVOICE_API_KEY" \
  -v omnivoice-data:/app/omnivoice_data \
  ghcr.io/debpalash/voicestudio:stable

How to use it

Each job has its own page in the sidebar. Pick the job, give it the input and generate.

A script or an agent can talk to the backend on port 3900. It speaks an OpenAI-style API, and loopback requests need no key. An MCP server sits on the same backend at http://localhost:3900/mcp/, and the app exports ready-made configs for Claude Code, Cursor and Codex CLI. The kit’s setup notes carry the maker’s request and config text.

Example

You want a short narration in your own voice, and you have a clean clip of yourself.

Voice cloning, reference recording: my-voice.wav (about 10 seconds, one speaker)
Text: Welcome back. Today we walk through the new pricing page.

The app saves a short audio clip in a voice meant to match your recording. The default engine writes 24 kHz mono audio. An invisible AudioSeal watermark that marks it as AI-made is on by default. The take lands in the generation history, and Settings, Privacy sets how many takes it keeps.

For newcomers and experts

If you are new. A cloning engine copies the sound of a short recording, so the quality of that recording matters. Use one speaker in a quiet room. Start with the demo voice that ships in the app before you try your own. A clean first result tells you the install works, and leaves your own recording as the only open question. Without a GPU, the Windows guide suggests a light engine such as KittenTTS, Supertonic-3 or PocketTTS, and a small Whisper model. Leave dubbing for later. It chains transcription, translation and speech, and the app has you review the translated text before it generates anything.

For experts. A FastAPI backend sits under the app, and the engine catalogue is wider than the default.

• Engines: OMNIVOICE_TTS_BACKEND and OMNIVOICE_ASR_BACKEND pin one. The guides run from light CPU engines to an 8B model
• Device: OMNIVOICE_DEVICE pins the device, and OMNIVOICE_TORCH_VARIANT=cuda, cpu or rocm overrides the PyTorch build
• Storage: OMNIVOICE_ENGINES_DIR and OMNIVOICE_DATA_DIR move engines and data off the system drive. A sidecar engine such as IndexTTS 2.5 gets its own Python environment and can add several GB

Good to know

The AGPL covers the app, not the voice models. Cloning and network use each have their own limits.

• Model licenses: the notice says the default k2-fsa/OmniVoice model card lists its code as Apache-2.0 and its pretrained weights as CC-BY-NC. Its audio tokenizer carries separate Boson Higgs Audio 2 and Meta Llama community terms. Other engines carry their own terms in the Model Manager, and the maker’s inventory records no model as cleared for commercial use. The site adds that buying Pro does not override a model’s terms
• Cloning a person: “Clone a voice only with its owner’s permission,” the README says. The app can record a spoken consent statement on a profile. Its own text says that check is for agentic features and community sharing, “never for local synthesis.” Nothing stops a local clone. Getting permission is on you, and laws on voice likeness differ by place
• Watermark: an invisible AudioSeal watermark marks new audio as AI-made. It is on by default, with a switch in Settings, Privacy. The maker’s privacy notes say that if the watermark model fails to load, audio is still made, without the mark
• AGPL-3.0 in plain words: you may use, copy, modify and redistribute the app, including for business. If you change it and let others use your version over a network, you must offer them its complete source under the same license
• Network: the install line reaches voicestudio.sh, and the installer and update check reach GitHub. Setup downloads Python packages, and weights come from Hugging Face or, when its automatic check finds it faster, the hf-mirror.com mirror. The maker says the updater has no telemetry and product analytics are opt-in behind a first-run prompt
• Translation: Argos and NLLB ship in every build and translate offline. Google, DeepL, MyMemory and Microsoft send the text online and need a package that, per the guide, only a from-source install can add. The LLM engine can point at a hosted service or at a local Ollama or LM Studio model
• Limits: Intel Macs run the UI only, Windows on ARM is experimental and CPU-only, and the phone apps are previews. The README’s 646 languages is its own claim, and the site says language availability varies by engine
• Pro: the maker’s site also lists a paid, closed-source Pro plan with a commercial license. This kit covers only the free app
• Name trap: PyPI has a package called voicestudio, but it is a different project. Install this one from the lines above, not with pip
• License: AGPL-3.0-only

If you used the install script on macOS or Linux, quit the app and run this line to remove it. Your data stays until you use Settings, Storage, Remove all data.

curl -fsSL https://voicestudio.sh/install | sh -s -- --uninstall