# Voicebox

> Open-source local-first AI voice studio for cloning voices, generating speech across 7 TTS engines, dictating into any app, and giving AI agents a voice.

Voicebox is a free, open-source AI voice studio built by Spacedrive Technology Inc. that runs entirely on your machine — no cloud required, no API keys, no per-character fees. It positions itself as a local alternative to both ElevenLabs (voice output) and WisprFlow (voice input), combining both halves of the voice I/O loop in a single desktop app. As of v0.5.0, the project has accumulated nearly 49,000 GitHub stars and the homepage reports over 1.9 million downloads.

## What It Is

Voicebox is a desktop application for voice cloning, text-to-speech generation, speech-to-text dictation, and AI agent voice integration. It bundles seven TTS engines, two Whisper-based transcription models, and a local Qwen3 LLM for transcript refinement and voice personality features — all running on the user's hardware via Metal (Apple Silicon), CUDA (NVIDIA), ROCm (AMD), DirectML (Windows), or CPU. The app is built with Tauri (Rust) for native performance rather than Electron, with a React/TypeScript frontend and a FastAPI Python backend.

## Core Capabilities

- **Voice cloning** — zero-shot cloning from as little as 3 seconds of reference audio, across seven engines: Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, and Kokoro
- **23 languages** — broadest coverage via Chatterbox Multilingual, including Arabic, Hindi, Japanese, Swahili, and more
- **Global dictation** — hold a hotkey anywhere on the OS, speak, release; transcript pastes into the focused text field on macOS with accessibility-verified injection
- **Stories editor** — multi-track timeline for composing multi-voice conversations, podcasts, and narratives
- **Audio effects pipeline** — pitch shift, reverb, delay, chorus, compression, and filters via Spotify's Pedalboard library, with per-profile presets
- **Unlimited generation length** — auto-chunking at sentence boundaries with crossfade, up to 50,000 characters per run
- **Voice personalities** — attach a free-form persona to any voice profile; a bundled Qwen3 LLM can Compose fresh lines or Rewrite input text in character

## MCP and Agent Integration

Voicebox ships a built-in Model Context Protocol (MCP) server, making it directly usable by MCP-aware agents such as Claude Code, Cursor, Cline, and Windsurf. A single tool call — `voicebox.speak` — lets any agent speak to the user in a cloned voice. Four MCP tools are exposed: `voicebox.speak`, `voicebox.transcribe`, `voicebox.list_captures`, and `voicebox.list_profiles`. The same endpoint is also reachable via a plain `POST /speak` REST call for non-MCP clients, shell scripts, or custom harnesses. Per-agent voice bindings let users assign different cloned voices to different agents (e.g., Claude Code in one voice, Cursor in another).

## Local REST API

Every downloaded TTS engine becomes a REST endpoint at `http://127.0.0.1:17493`. The OpenAPI reference is served at `/docs` when the app is running. Key endpoints include `/generate`, `/speak`, `/transcribe`, `/profiles`, `/history`, and `/health`. This makes Voicebox usable as a local voice backend for games, apps, scripts, and Stream Deck automations — with no rate limits and no external dependencies.

## Update: v0.5.0

The latest release is v0.5.0, published April 25, 2026. The GitHub repository was created January 25, 2026 and last pushed July 28, 2026, indicating active development. The roadmap lists upcoming features including Windows/Linux auto-paste parity, additional STT engines (Parakeet v3, Qwen3-ASR), streaming transcription via WebSocket, end-to-end speech LLMs (Moshi, GLM-4-Voice, Qwen2.5 Omni), voice design from text descriptions, long-form dual-stream capture, and a mobile companion app. The project is licensed under MIT and the homepage states it is "open source, runs entirely on your machine."

## Features
- Voice cloning from as little as 3 seconds of audio
- 7 TTS engines: Qwen3-TTS, Qwen CustomVoice, LuxTTS, Chatterbox Multilingual, Chatterbox Turbo, HumeAI TADA, Kokoro
- 23 language support
- Global dictation hotkey with push-to-talk and toggle modes
- Whisper-based speech-to-text (Base, Small, Medium, Large, Turbo)
- Stories editor for multi-voice timeline composition
- Audio effects pipeline (pitch shift, reverb, delay, chorus, compression, filters)
- Unlimited generation length up to 50,000 characters with auto-chunking
- Voice personalities with Compose and Rewrite via local Qwen3 LLM
- Built-in MCP server for agent voice integration
- REST API at localhost with no rate limits or API keys
- Per-agent voice binding for MCP clients
- Captures tab with original audio and transcript storage
- Paralinguistic tags support ([laugh], [sigh], [gasp]) via Chatterbox Turbo
- Delivery instructions via natural language for Qwen3-TTS
- GPU acceleration: Metal, CUDA, ROCm, DirectML, Intel Arc, CPU
- On-screen pill overlay for dictation and agent speech states
- LLM transcript refinement for ums and stutters
- Multi-sample voice profile support
- Import/export voice profiles
- Generation version history with source tracking
- Async non-blocking generation queue
- Docker support
- Built with Tauri (Rust) for native performance

## Integrations
Claude Code, Cursor, Cline, Windsurf, VS Code MCP extensions, OpenAI Whisper, Qwen3-TTS, Chatterbox (Resemble AI), HumeAI TADA, Kokoro, LuxTTS, Spotify Pedalboard, MLX (Apple Silicon), PyTorch (CUDA/ROCm/CPU), FastAPI, SQLite, WaveSurfer.js, Docker

## Platforms
WINDOWS, MACOS, LINUX, API, VSC_EXTENSION, CLI

## Pricing
Open Source, Free tier available

## Version
v0.5.0

## Links
- Website: https://voicebox.sh
- Documentation: https://docs.voicebox.sh
- Repository: https://github.com/jamiepine/voicebox
- EveryDev.ai: https://www.everydev.ai/tools/voicebox