VoiceStudio

What it does

Six jobs, one local app.

Start with the default VoiceStudio engine, which the project powers with OmniVoice. The Model Manager can add other engines later, and it shows each model’s own licence. The project describes language coverage as 646 languages. Treat that as the project’s claim and test the language you actually need.

• Create

  Voice cloning

  A clean reference recording becomes speech in that voice. Use a recording only when the speaker has allowed it.

• Create

  Voice design

  Describe a voice in words. The app maps that description to traits such as age, pitch and style, then saves a design profile.

• Create

  Video dubbing

  Translate speech and time it to the original video. The docs include a multi-speaker dubbing guide and an export guide.

• Capture

  Dictate and transcribe

  A floating dictation widget types into the focused field. Transcription can also run on a file, including live engines such as sherpa-onnx.

• Produce

  Narrate and batch

  Stories, audiobooks and a queue of jobs. Useful when one short test sounds right and you want the same voice on a longer script.

• Connect

  API, MCP and recipes

  Other programs on the same computer can call the local speech service, or an agent can use the MCP tools while the app is open.

Where it runs

Install it where you will actually listen.

Desktop and phone installers are on the project’s latest GitHub release. Docker images come from GHCR. No GPU is required. CPU mode is slower. Native GGUF engines can also use Vulkan on AMD, Intel and NVIDIA GPUs.

macOS Apple Silicon

Stable .dmg. Metal acceleration. If Gatekeeper says the app is damaged, that is a quarantine block, not a broken download. The macOS install doc explains the fix.

Windows 10/11 x64

Stable .exe. NVIDIA CUDA, optional AMD ROCm, or CPU.

Linux x64

Stable AppImage, .deb or Nix. NVIDIA CUDA or CPU.

Docker

Stable image. NVIDIA CUDA or AMD ROCm.

Android 10+

Preview .apk. On-device and offline.

iPhone / iPad

Preview for iOS 17+, installed by sideload. On-device and offline.

Windows on ARM

Experimental .exe. CPU through emulation.

Intel Mac

Limited. The app UI can connect to a remote backend rather than running the full local stack.

How a first session works

Generate one short line before you automate anything.

• 01

  Install a build you can identify

  Prefer the latest release installer for your system. The project also documents a shell installer, a PowerShell installer, and a from-source path with bun run dev. Read the install doc for the path you pick.

• 02

  Let the model download be a choice

  Open Voice cloning. The app asks before it downloads a model. Accept one model for the test, not the whole catalogue.

• 03

  Use a voice you have rights to

  Pick a built-in voice or add a clean reference recording. Type one sentence and generate. Judge clarity, timing and how hot the machine gets.

• 04

  Change engine only if the default fails the job

  Speech options in the catalog include CosyVoice 3, IndexTTS, KittenTTS, gpt-sovits, sherpa-onnx, MOSS-TTS and GGUF builds. Transcription defaults to WhisperX, with Faster-Whisper, Parakeet, Moonshine and FunASR among the others. Availability depends on hardware and what you installed.

How other tools use it

The app is also a local speech service.

While VoiceStudio is open, the backend listens on port 3900 unless you move it. Loopback clients do not need a key. A remote caller cannot turn on the desktop microphone. Sharing beyond your computer uses a PIN and belongs on a trusted network, not the public internet.

Dictation

Use the in-app shortcut to insert text into the focused field.

Discover

GET http://127.0.0.1:3900/.well-known/voicestudio-speech describes the local speech API.

Transcribe a file

POST http://127.0.0.1:3900/v1/audio/transcriptions is the OpenAI-compatible HTTP path.

Live transcription

WebSocket at ws://127.0.0.1:3900/v1/audio/transcriptions/stream for partial and final text.

Agents

MCP is at http://127.0.0.1:3900/mcp/ with the trailing slash. Tools cover speech, cloning, voice design, transcription and listing voices. The desktop Integrations screen exports configs for Claude Code, Cursor and Codex.

Coding agents

The repository documents npx skills add debpalash/VoiceStudio for a voicestudio skill and a maintainer skill. Those skill files stay in the VoiceStudio repo.