Clone your voice for AI agents and type by speaking in any app with a local studio
Hold a microphone instead of a keyboard, and your speech is converted to text and pasted where you want it. Voicebox bundles speech recognition and synthesis into a single app that handles all processing on your computer. Without relying on the cloud, you can clone your voice or experience AI agents answering in your voice.
Voice input and output without the cloud
Voicebox serves as an alternative to cloud services like ElevenLabs or WisprFlow, handling both speech synthesis and recognition locally. Models, voice data, and recordings never leave your device, preserving privacy. The app is built with Tauri (Rust), making it lighter than Electron, and utilizes platform-specific GPUs such as MLX on macOS or CUDA on Windows. This allows for fast processing speeds without needing to connect to external servers.
Choosing a voice from eight engines
It offers eight TTS engines, including Qwen3-TTS, Chatterbox, and Kokoro, allowing you to switch engines for each generation. With just a few seconds of audio, you can clone a voice via zero-shot cloning or use one of over 50 preset voices. It supports 23 languages, and using Chatterbox Turbo allows you to express emotions with tags like [laugh] or [sigh]. It also includes Qwen VoiceDesign, which creates new voices from text descriptions alone.
Giving your voice to agents
AI agents that support MCP (Model Context Protocol) can speak in your voice. A single voicebox_speak tool call lets agents like Claude Code or Cursor deliver task completion notifications or questions in your cloned voice. You can assign specific voices to each agent so you immediately know who is speaking. Additionally, you can use a local LLM to assign a persona to a voice and have the agent rewrite its responses to match that persona.
Installation and requirements
It is distributed for macOS (Apple Silicon/Intel), Windows, and Docker, while Linux requires building from source. macOS requires Accessibility permissions, and Windows and Linux benefit from CUDA or ROCm environments for GPU acceleration. To clone for development, you need Python 3.12, Rust, Bun, and Tauri Prerequisites. You can set up the environment with the just setup command and run it with just dev. It is released under the MIT license, allowing free use.
Limitations and checks for local environments
Since all processing happens locally, low GPU performance may result in slower processing speeds. Automatic pasting is not yet available on Linux, while automatic pasting for Windows and Linux is on the roadmap. Issues may arise with model downloads and GPU configuration, so it is advisable to check the troubleshooting guide in advance. Unlike cloud services, it works without an internet connection, but a network is required for the initial model download.
