Voicebox: The Open-Source AI Voice Studio for Cloning and Dictation

TL;DR: jamiepine/voicebox is a TypeScript open-source AI voice studio that bundles voice cloning, dictation, and creation into one app. It tops 54,000 stars and supports CUDA, MLX, Whisper, and Qwen3-TTS.

What it is

jamiepine/voicebox describes itself as “The open-source AI voice studio. Clone, dictate, create.” The repo is by jamiepine, the language is TypeScript, the license is MIT, and it sits at 54,416 stars with 6,791 forks. It is not archived.

Topics include ai, cuda, mlx, qwen3-tts, qwen3-tts-ui, voice-ai, voice-clone, and whisper, which is a fairly complete map of the modern open-source TTS stack. The homepage is https://voicebox.sh.

What it does well

A voice studio that bundles three workflows — cloning, dictation, and creation — in one TypeScript app is unusual; most open-source projects pick one. Voicebox covers all three.

• Voice cloning. You can train a clone from a short reference clip and then speak with that voice in the studio, leveraging Qwen3-TTS or Whisper-derived backends.
• Dictation. The transcription pipeline uses Whisper for accurate speech-to-text, which is the de facto open-source ASR baseline.
• Creation. The “create” path lets you drive new TTS outputs from text prompts, with the choice of CUDA on NVIDIA GPUs or MLX on Apple Silicon so the studio runs natively on both ecosystems.
• Web UI. A TypeScript app means the same code base ships to a local web UI, which makes it accessible from any machine on the LAN without a separate Electron build.

How to try it

The repo ships its install instructions; for a self-hosted install on Apple Silicon the path looks roughly like:

git clone https://github.com/jamiepine/voicebox.git
cd voicebox
npm install
npm run dev   # local web UI

The CUDA path follows the same shape with the GPU-backed Python backend wired in for inference.

When it fits, and when it doesn’t

Voicebox fits anyone who wants a self-hosted alternative to hosted voice cloning APIs and is comfortable installing a Node-based app and a Python inference backend. It is less attractive if you want a turnkey, no-install voice product — a hosted service will be faster to first audio — or if you only need a single workflow, since several single-purpose tools in this category have smaller code surfaces to maintain.

54k stars and a topic list that includes both CUDA and MLX are useful signals that the project has been kept current with the underlying model ecosystem.

Source and Accuracy Notes

• Project page: https://github.com/jamiepine/voicebox
• Source repository: jamiepine/voicebox (verified via GitHub API html_url)
• License: MIT (verified via GitHub API license.spdx_id)
• Stars at time of writing: 54,416
• Language: TypeScript
• Homepage: https://voicebox.sh
• Topics: ai, cuda, mlx, qwen3-tts, qwen3-tts-ui, voice-ai, voice-clone, whisper