Voicebox clones a voice entirely on your own machine
Your voice can be copied now. This one does it on the machine in front of you, with nothing sent anywhere. voicebox clones a voice, types what you say, and gives your coding agent a voice to answer in. Free, MIT-licensed, completely offline.
It's a desktop voice studio.
It's a desktop voice studio. You clone a voice from a sample, generate speech in that voice, and hold one hotkey to dictate into any app you're already in. The project pitches it as a free, local alternative to ElevenLabs and Wispr Flow in one app. Its argument: one cloud owns the talking, the other owns the typing, and voicebox runs both on your machine.
Cloning is almost no ceremony.
Cloning is almost no ceremony. Hand it ten to thirty seconds of clear speech, tell it the exact words you spoke, and you have a profile. And the project says the quiet part in the app itself: it does not and cannot check whose voice you handed it. Then it lets you decide.
Down the loop: your voice becomes clean text.
Here's how the typing half works. You hold a chord anywhere on the system. Your voice goes to Whisper. Out comes raw text. That text goes to a small local model that strips the ums and false starts. Then it pastes the clean text for you. And this part's clever: it pastes into the field that had focus when you started the chord, not wherever it drifted while you talked.
Now run it backwards.
Now run it backwards. You hand it text, and the same local model rewrites it. Not to clean it up this time, but to say it the way a character would. Then it goes out through one of the speech engines. Here's the idea. The model that decides you didn't mean to say um is the same model that decides how a character would have said it. One local model doing both jobs, not two. The project's own line is on the slide.
Your agent gets a voice.
Now hand that loop to your coding agent. voicebox runs a small MCP server, so Claude Code or Cursor picks up a voice from one tool call. Bind each agent to its own voice, and you can hear which one is talking. And it can never speak behind your back. Every agent voice raises the same pill your dictation does. The project's reason: silent background speech is a trust hazard.
Seven engines, and they're not interchangeable.
Here's what the headlines get wrong. voicebox isn't one voice model with a UI on it. It's a switchboard over seven, and you pick one per job. That "23 languages" number? That's a single engine, Chatterbox Multilingual. The rest are ten languages or fewer, and two are English only. And only Chatterbox Turbo reads tags like laugh and sigh. On the Qwen engine you steer delivery in plain words: speak slowly, warm and friendly.
Get it.
Getting it. It's MIT, built on Tauri and Rust with a Python backend, not another Electron app. It runs on Apple Silicon, NVIDIA, AMD, Intel, or plain CPU. One caveat: there's no prebuilt Linux binary yet, so on Linux you build from source. On a Mac the transcript pastes right into the field you're in. Latest version is 0.5.0.
The whole loop stays on your desk.
So if you assumed that copying a voice means handing it to somebody's cloud, this is the version where the whole loop stays on your desk. Clone, dictate, and give your agent a voice, all offline. The link is on screen. New repo tomorrow.
