Voicebox Setup: A Local AI Voice Studio in 15 Minutes

Voicebox is a free, open source AI voice studio that clones voices, generates speech in 23 languages, and handles system wide dictation entirely on your own machine, and it is one of today's fastest climbing repositories on GitHub, adding roughly 610 stars today on its way to about 43,600 total. It is worth setting up because it collapses two separate paid cloud subscriptions, one for text to speech and one for dictation, into a single MIT licensed desktop app. Budget about 15 minutes: the installer is a large download, and the models fetch on first launch.

• Voicebox v0.5.0 ships ready made installers for Windows and macOS, plus a Docker path; Linux has no prebuilt binary yet and builds from source.
• Seven TTS engines are switchable per generation, including Qwen3-TTS, Chatterbox Multilingual and Kokoro, with zero shot voice cloning from a short reference sample.
• A built in MCP server lets Claude Code, Cursor or Windsurf speak to you in a voice you cloned, through one voicebox.speak tool call.
• Everything is local by default: TTS, Whisper transcription and a bundled Qwen3 LLM all run on your GPU, and it is built on Tauri rather than Electron.

What is Voicebox and why is it trending?

Voicebox describes itself as a local first AI voice studio, and a free alternative to the two cloud incumbents that between them own the voice loop: one sells text to speech and voice cloning, the other sells dictation. The project's argument is that these are two halves of the same problem, and that both halves now fit on a laptop. It is a Tauri app, so the shell is Rust rather than Electron, with a React and TypeScript frontend, a FastAPI Python backend, and SQLite for storage. Audio effects come from Spotify's pedalboard library.

RelatedMeetily Setup: Free Local AI Meeting Notes in 10 Minutes

The engine list is the unusual part. Rather than betting on one model, Voicebox bundles seven and lets you switch per generation: Qwen3-TTS and Qwen CustomVoice for high quality multilingual cloning, LuxTTS as a lightweight option the README puts at roughly 1GB of VRAM and 150x realtime on CPU, Chatterbox Multilingual for the broadest language coverage at 23 languages, Chatterbox Turbo for fast English with emotion tags, HumeAI TADA for long form coherence, and Kokoro as a tiny 82M model with 50 preset voices. Transcription runs OpenAI's Whisper, on MLX for Apple Silicon or PyTorch everywhere else. The trending spike tracks the obvious pressure point: cloud voice generation is metered per character, and cloned voice data is about as personal as data gets.

How do you install Voicebox on Windows, macOS and Docker?

Grab the installer for your platform from the v0.5.0 release. On Windows that is Voicebox_0.5.0_x64-setup.exe at about 517MB, or the equivalent MSI at about 518MB. On macOS, Apple Silicon takes Voicebox_0.5.0_aarch64.dmg at roughly 513MB and Intel takes Voicebox_0.5.0_x64.dmg at roughly 560MB; open it and drag Voicebox into Applications. Nvidia users on Windows should know the app auto downloads a CUDA server binary from inside the app, and that asset is a further 987MB, with a CUDA libraries archive of about 2GB on top. Plan disk accordingly.

Docker is a one liner:

# from the repo root
$ docker compose up

Linux has no packaged build today, so you compile it. The repo uses just as its task runner, and the prerequisites are Bun, Rust, Python 3.11 or newer, the Tauri prerequisites, and Xcode on macOS:

# build from source
$ git clone https://github.com/jamiepine/voicebox.git
$ cd voicebox
$ just setup   # creates Python venv, installs all deps
$ just dev     # starts backend + desktop app

Install the just task runner itself via Homebrew (brew) or Cargo, and run just --list to see every command. On first launch, pick a TTS engine and a Whisper size; the models download once and then run offline. GPU support is broad: MLX and Metal on Apple Silicon, CUDA on Nvidia, ROCm on AMD, DirectML for any Windows GPU, IPEX for Intel Arc, and CPU everywhere as the slow fallback.

How do you give Claude Code or Cursor a voice?

This is the feature most people will actually keep. Voicebox mounts an MCP server on its local API port, so any MCP aware agent can speak through a voice you cloned. For Claude Code it is one command:

$ claude mcp add voicebox --transport http --url http://127.0.0.1:17493/mcp --header "X-Voicebox-Client-Id: claude-code"

Any other HTTP MCP client, Cursor and Windsurf included, takes the same endpoint as JSON:

{
  "mcpServers": {
    "voicebox": {
      "url": "http://127.0.0.1:17493/mcp",
      "headers": { "X-Voicebox-Client-Id": "cursor" }
    }
  }
}

Four tools ship: voicebox.speak, voicebox.transcribe, voicebox.list_captures and voicebox.list_profiles. You can pin a different voice per client in Settings, so Claude Code and Cursor sound distinct and you know which agent is talking without looking. Anything that does not speak MCP can hit the REST API directly:

# Agent voice output from any script
$ curl -X POST http://127.0.0.1:17493/speak \
    -H "Content-Type: application/json" \
    -H "X-Voicebox-Client-Id: my-script" \
    -d '{"text": "Deploy complete.", "profile": "Morgan"}'

Full API docs are served locally at http://127.0.0.1:17493/docs.

RelatedVibe-Trading Setup: Self-Host an AI Trading Agent Free

The cloud products still hold real advantages: no multi hundred megabyte download, no GPU requirement, no model management, and voice quality that does not depend on your hardware. Voicebox wins wherever recurring cost or data control is the binding constraint rather than a preference.

What are the gotchas before you rely on it?

Five things worth knowing before this becomes part of your daily workflow. First, size: the desktop app alone is over half a gigabyte, and an Nvidia setup pulls down roughly 3GB more in CUDA server and library assets. This is not a quick trial on a thin laptop or a metered connection. Second, dictation auto paste is macOS only today. The README lists accessibility verified injection into the focused text field as a macOS feature, and puts Windows and Linux paste parity on the roadmap rather than in the current build, so Windows users get the capture and the transcript without the seamless paste.

Third, emotion tags are engine specific and this trips people up: only Chatterbox Turbo interprets paralinguistic tags like [laugh] and [sigh]. Every other engine, Qwen3-TTS and LuxTTS included, reads them out literally as text. Fourth, Linux is a source build, not an installer, and it wants Bun, Rust, Python and the full Tauri toolchain in place first. Fifth, and least technical but most important: zero shot cloning from a few seconds of audio is exactly as consequential as it sounds. Cloning a voice that is not yours, or that you do not have explicit permission to use, is the kind of thing that ends in a legal letter regardless of where the model runs.

Our take

The interesting claim here is not that local TTS exists, because it has for a while. It is the packaging. Seven engines, Whisper, a bundled Qwen3 LLM for refinement and personas, and an MCP server all sharing one model cache and one GPU memory footprint, behind an installer a non engineer can double click, is a genuinely difficult integration job and it is the part nobody had done. The MCP angle is the sleeper feature: giving a coding agent a voice sounds like a toy until you have spent a long build waiting on a terminal, and a sentence spoken in a voice you picked turns out to beat a notification badge.

The honest caveats are size and platform parity. This is a multi gigabyte install that really wants a GPU, and the dictation half, arguably the half you would use hourly, is still noticeably better on macOS than on Windows. If you are on Apple Silicon, this is an easy recommendation today and it replaces two subscriptions. If you are on Windows without an Nvidia card, or on Linux, it is worth watching for a release or two while the roadmap items land. Either way, the direction is clear: the voice layer is following the same path the transcription layer already walked, from metered cloud API to something that just runs on your machine.

Original analysis by GenZTech. Tool documentation: jamiepine/voicebox on GitHub.