Voicebox

Voicebox builds a voice profile from a short sample instead of a long training session. You can drop in an audio file, record up to 30 seconds from your microphone, or capture audio playing on your system, which means you can clone a voice straight from a video or podcast. Profiles are reusable. You can combine several samples to sharpen the result. For anyone who wants open source voice cloning without renting cloud minutes, this is the core of the tool.

Rather than betting on one model, Voicebox bundles seven text-to-speech engines and lets you switch between them. Qwen3-TTS handles multilingual cloning, Chatterbox Multilingual covers the widest language range, Chatterbox Turbo adds paralinguistic tags like laughs and sighs, and Kokoro stays tiny and fast on CPU. There isn't a single best engine, just the right one for a task. Picking per job beats being locked to whatever a single vendor ships.

A global shortcut turns your speech into text in any app. Hold the keys, talk, and release; the transcript drops into the focused field or your clipboard. It's the part that lets you dictate into any app on macOS and Windows, and every text box gets an in-app microphone button. The transcript and the original audio are kept together in Captures, so you can revisit what you said later. It's a practical answer to typing long messages by hand.

One tool call, voicebox.speak, lets any MCP-aware agent talk in a voice you've cloned. Claude Code, Cursor, Cline, and anything else that speaks MCP can bind to a voice profile, so you hear which agent is responding without looking. Every agent-initiated clip surfaces the same on-screen pill, so nothing plays silently in the background. The endpoint also fits other tool-call protocols, like ACP and A2A, on the same local server.

Every engine you download becomes a REST endpoint on your own machine at port 17493. There are no API keys, no rate limits, and no per-character fees, because the requests never leave your computer. Developers use it to generate NPC dialogue, localize characters, or add voice replies to an app, all against a localhost address. For anyone building voice features, it removes the usual pricing and privacy trade-off.

A voice profile can carry a free-form personality, and the app's local model will rewrite your text in that voice or improvise a fresh line on demand. That's handy for game dialogue, narration cues, or keeping a character consistent across a long project. The Stories editor then lays multiple voices on a multi-track timeline, so you can cut a scene or a podcast segment with several speakers. It turns scattered clips into a finished piece. No studio required.