Run your own voice studio on your own machine: hands-on with VoiceStudio
- Open-source under AGPL-3.0; 53,900+ stars and 6,000+ forks; #1 trending repository last week with 16,700+ stars in seven days; powered by default by the k2-fsa/OmniVoice engine.
- No API keys — everything runs on your hardware; local speech service on port 3900 (loopback) with OpenAI-compatible endpoints plus an MCP server at /mcp that Claude Code and Cursor can call.
- Install: curl -fsSL https://voicestudio.sh/install | sh (macOS/Linux) or irm https://voicestudio.sh/install | iex (PowerShell); version 0.5.3 was the final Tauri release — the Electron app is now the only desktop app.
- GPU optional (NVIDIA CUDA, Apple Silicon Metal; CPU-only works, slower); the README’s license section is explicit: clone voices only with permission, and review model licenses before commercial use.

Every voice API bills you per character. The hosted voice platforms charge for synthesis the way cloud GPUs charge for compute — small amounts feel free, real workloads get expensive, and your audio runs through someone else's servers forever.
VoiceStudio bets that voice work shouldn't be a cloud service at all. It's a fully local, open-source voice studio: clone voices, design new ones from a text description, dub videos with timed speech, dictate, transcribe, and produce audiobooks — in 646 languages, powered by default by the k2-fsa/OmniVoice engine. It was the #1 trending repository on GitHub last week, adding more than 16,700 stars in seven days. Today it stands at 53,900+ stars and 6,000+ forks, released under AGPL-3.0.
What you get at the end: a desktop app where generating a voice costs nothing, plus a local speech service your coding agents can call over HTTP and MCP. Here's the whole thing, hands-on.
What you'll need
- A Mac, Windows PC, or Linux machine. NVIDIA GPUs get CUDA acceleration and Apple Silicon uses Metal; machines with no dedicated GPU are fully usable on the CPU, just slower. The installer pulls a small CPU build of PyTorch (about 5 GB of free disk) when you have no GPU.
- Free disk for model downloads. The app installs models on demand, the first time you use each engine.
- No API keys. Everything runs on your hardware; remote services are optional.
1. Install with one command
On macOS or Linux, one command installs the Electron desktop app:
curl -fsSL https://voicestudio.sh/install | sh
On Windows, run this in PowerShell:
irm https://voicestudio.sh/install | iex
Prefer a manual download? Grab macOS .dmg, Windows .exe, or Linux .AppImage / .deb from the Releases page and follow the platform guides in docs/install/. The installer preserves your settings, projects, and models across upgrades. One caveat from the README: version 0.5.3 was the final Tauri release — the Electron app is now the only desktop app.
If you live inside a coding agent, you can delegate the whole install. Paste this into Claude Code, Codex, or Cursor:
Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md
The agent guide covers hardware detection, reusing existing data, asking before model downloads, and a test generation. Agents that support skills can also run npx skills add debpalash/VoiceStudio to get the voicestudio audio-workflow skill.
2. Clone your first voice
Open the Voice cloning workspace, pick a voice or add a clean reference recording, enter your text, and generate. The app installs the required model when prompted. That is the README's entire quick-start, and it's honest: the workflow is genuinely that short.

Two things worth knowing before you record. First, a clean reference beats a long one — a quiet room with no background noise produces a better clone than a long noisy recording. Second, the README's license section is explicit: clone voices only with permission. The same model licenses that enable cloning require you to review them before commercial use.
3. Design a voice from a sentence
No reference recording? The Voice design workspace generates a voice from a plain-language description — gender, age, pitch, style, accent — and gives it a stable identity you can reuse. This is also where the MCP tools shine: describe_voice tells you which attributes your description maps to (and which words it couldn't map), while design_voice mints a new voice profile and returns a profile_id you can generate with later. It refuses descriptions that map to no attribute, which is a polite way of saying "be more specific."
4. Transcribe and dictate
The transcription workspace handles audio files in 646 languages. Dictation is a floating widget with a global shortcut — the app owns microphone capture, and the Python backend keeps ASR models warm.
Scripting transcription without the UI is one command, using the dependency-free Python bridge that ships with the app:
python -m backend.speech_client capabilities
python -m backend.speech_client transcribe recording.wav
It targets VOICESTUDIO_URL, else the loopback backend on port 3900, and sends OMNIVOICE_API_KEY as a Bearer POST :3900/v1/audio/transcriptions speaks the same contract — or stream live partial text over the WebSocket at :3900/v1/audio/transcriptions/stream. A machine-readable capability document lives at http://127.0.0.1:3900/.well-known/voicestudio-speech.
5. Dub a video
The Dubbing workspace turns timed speech into translated video: upload a video, generate speech in another language, and get timed dubbing out. It's the same engine stack as cloning, aimed at the most tedious part of video localization — matching translated speech to the timeline.

Pair this with the audiobook and batch-job workflows if you're producing content at volume: generate hours of narration once, pay nothing, and never send the source material to a third party.
6. Give your coding agent a voice
VoiceStudio ships an MCP server so agents like Claude Code and Cursor can synthesize speech, clone voices, and transcribe — in a voice you choose per agent. The server is mounted on the running backend at /mcp, so there's nothing extra to start once VoiceStudio is open. For stdio-only clients, the shim is:
python -m backend.mcp_shim
The tools, from docs/mcp.md:
generate_speech— text to WAV by default (Opus/Ogg when ffmpeg is installed), in the agent's bound voice or anyprofile_id.clone_voice/design_voice— reference audio or a text description in, a newprofile_idout.describe_voice— dry-run of a description: which attributes it maps to, which words it didn't understand. Saves nothing.transcribe— audio in, text out, 646 languages.list_voices/list_personalities/list_languages— enumerate what's available.check_health— backend status plus the active GPU device.
One practical detail: a WAV as base64 is a lot of bytes in an agent's context. Setting OMNIVOICE_MCP_OUTPUT_MODE=files (or both) makes generate_speech write the audio to disk and hand back an audio_url served at /audio/<audio_id>.<format> plus an output_path — the agent gets a file path instead of megabytes of base64.
What you built
A voice studio with no meter running: a cloned voice and a designed voice saved to your library, a transcription pipeline you can hit from shell scripts or an OpenAI-compatible endpoint, a dubbed video, and a coding agent that can narrate its own work in any voice you choose. Per character billed: zero.
Honest limitations
- AGPL-3.0. Great for personal and internal use; if you distribute the app or offer it as a hosted service, the copyleft applies. Models ship under their own licenses — review them before commercial use.
- First-run downloads. Nothing synthesizes until the engine's models are on disk, and they're large. Budget time and disk space; the installer's agent guide at least asks before downloading.
- CPU works, GPU flies. CUDA and Apple Silicon Metal are the fast paths; CPU-only machines are fully supported but slower.
- Platform gaps. Intel Macs get the app UI only and need a remote backend; Windows on ARM is experimental (native ARM64 app, emulated x64 Python backend, CPU only).
- Consent is on you. Cloning voices you don't own or have permission to clone is a misuse of the tool — the project says so explicitly in its license section.