jamiepine/voicebox

What it does

Voicebox packages voice cloning, text-to-speech, speech-to-text, and system-wide dictation into one desktop application. It is aimed at creators producing narration or podcast audio, developers adding voice I/O to agents, and users who want voice processing to stay on their own machine rather than a cloud service. The app combines 8 TTS engines, including Qwen3-TTS, LuxTTS, Chatterbox, HumeAI TADA, and Kokoro, with Whisper for transcription, and adds a bundled local Qwen3 LLM for transcript cleanup and persona-based rewriting. Voice cloning works zero-shot from a short reference sample, and each engine exposes different strengths: language coverage, preset voices, expressive tags, or CPU speed Overview Engine table.

Beyond single clips, the app offers an auto-chunking pipeline that crossfades long scripts, a multi-track Stories editor for dialogue and podcasts, generation versioning with effect chains built on Spotify's Pedalboard, and an async queue that serializes GPU work Long-form generation Versions Effects. A global dictation hotkey, MCP server, and REST API (/generate, /speak, /transcribe) let other applications and agents use the same voices Dictation API.

Quick start

Install a prebuilt desktop build: macOS Apple Silicon and Intel DMGs and a Windows MSI are available from the download page, with all binaries listed on the releases page Download options Latest release. Linux has no prebuilt binary yet; use the Linux install instructions to build from source.

After launching, create a voice profile by uploading an audio sample or recording in the app, select a TTS engine and language, then generate speech. The same profile can be used for /speak and MCP-based agent output once the local backend is running.

To develop or build the app yourself:

git clone https://github.com/jamiepine/voicebox.git
cd voicebox
just setup   # creates Python venv, installs dependencies
just dev     # starts backend + desktop app

Prerequisites are Bun, Rust, Python 3.12, the Tauri prerequisites, and Xcode on macOS; install just first Development setup. For running as a container, the README lists docker compose up as the Docker path.

Requirements and limits

GPU support determines practical speed: MLX/Metal on Apple Silicon, CUDA on Windows and Linux NVIDIA, ROCm on AMD, DirectML on Windows, and Intel XPU. CPU inference works everywhere but is slower GPU table. Models are downloaded into a configurable directory (VOICEBOX_MODELS_DIR), and individual models can be unloaded to free GPU memory Model management.

Dictation auto-paste is described as accessibility-verified on macOS only; Windows and Linux paste parity is still on the roadmap Dictation Roadmap. Only Chatterbox Turbo interprets expressive tags such as [laugh]; other engines read them as literal text Emotions. Linux builds must be compiled locally. The project is MIT-licensed License.

HotRepo's take

The repository has recorded daily trends since June 2026, rising from roughly 31.9k stars to 59.8k by November 2026. During that period it reported daily gains of over 1,000 stars on 23–24 June, and on 7 November it reached 8th place in the Python daily list. Those figures show sustained attention rather than a single spike, but they are follower counts and list positions, not evidence of voice quality or user adoption.

What stands out technically is the breadth of the local stack: eight TTS engines plus Whisper STT, a bundled LLM sharing the inference runtime, and both REST and MCP interfaces around the same voice profiles. The README is explicit about platform gaps, especially Linux packaging and non-macOS paste, which are worth checking against your environment before committing to it.