Clone your voice for free with Voicebox and Claude
Hey, it's Cindy 🌱 Voicebox is an open source voice studio that clones your voice, generates speech in 23 languages, and gives you a dictation hotkey for any app on your computer.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
what it is
The repo
Voicebox lives at github.com/jamiepine/voicebox. It is MIT licensed and free, and it bundles three things people normally pay two separate subscriptions for: voice cloning, text to speech, and system-wide dictation. The whole thing runs locally, and the repo's own line on it is that "models, voice data, and captures never leave your machine."
Clone your voice Give it a reference recording and it builds a voice profile you can generate unlimited speech with, in your own voice.
Dictate anywhere A global hotkey types straight into whatever text field you are focused on, so you talk instead of type in any app.
23 languages Across seven engines, including Arabic, Hindi, Japanese, Swahili, Turkish and Polish, so your cloned voice can speak languages you do not.
It talks to Claude A built-in MCP server means Claude Code can speak out loud, transcribe audio, and browse your voice profiles. More on that below.
be honest with yourself
Which path is actually yours
The reel makes it look like one prompt does everything. Here is the true version, because it changes what you should do next.
On Mac or Windows, you do not need the terminal at all. There are prebuilt installers, so downloading one is genuinely the fastest route. Claude Code earns its keep on the parts after the install: wiring up the MCP connection, building from source on Linux, and untangling errors when a model download fails or a hotkey does not fire. Use the easy path first, and bring Claude in when something goes sideways.
Mac and Windows Download the installer, open it, done. No terminal, no Claude needed for this step.
Linux, or you want to build it yourself There are no prebuilt Linux binaries yet, so you compile from source. This is where Claude Code genuinely saves you an afternoon.
🔓 Open it and allow it. On macOS the first launch may need you to open System Settings, go to Privacy & Security, and click Open Anyway. Voicebox also asks for Accessibility and Microphone permissions, which is what lets the dictation hotkey type into other apps. Grant both or dictation will look broken.
⬇️ Let it pull a model. The speech engines download on first use, so the very first generation is slower than the rest. Stay on wifi for this part and let it finish before you judge the speed.
🎙 Record your reference clip. Go to voice profiles and record yourself reading something natural for a while. The docs do not state a minimum, so my rule is 20 to 30 seconds of clean audio, no music, no background noise, in the tone you actually want back. A rushed five second clip is the single most common reason a clone sounds off.
🔊 Generate your first line. Type a sentence, pick your profile, and generate. Compare it against the original recording, then re-record the reference with a bit more energy if it sounds flat. Two or three reference takes gets you most of the way.
⌨️ Turn on the dictation hotkey. Set your chord in the in-app picker. It supports hold-to-speak and tap-to-toggle, and on macOS it injects text into the focused field and restores your clipboard afterwards, so it does not wipe whatever you had copied.
If you have never used dictation properly, this is the feature that actually changes your day. Speaking a long message or a first draft is roughly three times faster than typing it, and the editing pass afterwards is usually shorter than you expect.
pick your engine
The seven engines, and when to use each
Voicebox ships several speech engines rather than one, which looks confusing until you know what each is for. This is the short version so you stop guessing.
Qwen3-TTS The default workhorse for cloning. High quality, multilingual, and it takes delivery instructions like "speak slowly" or "whisper".
Qwen CustomVoice Nine preset voices with natural-language delivery control and no reference audio needed. Good when you do not want your own voice on it.
Chatterbox Multilingual The broadest language coverage of the set. Reach for this one when the language matters more than the polish.
Chatterbox Turbo A fast 350M model with emotion and sound tags. Nice for quick drafts and playful reads.
LuxTTS Lightweight, around 1GB of VRAM, 48kHz output and very fast on CPU. This is your pick on an older or lower-spec machine.
Kokoro and TADA Kokoro is a tiny 82M model with 50 preset voices for fast CPU work. TADA (from Hume) holds coherence over 700 seconds plus, so it suits long-form reads.
copy this
The exact Claude prompt
This is the one from the reel. Use it when the download path does not apply to you, or when something breaks. Paste it into Claude Code with the repo open in front of you.
👩💻 install voicebox with me
I want to install Voicebox from https://github.com/jamiepine/voicebox on my machine. I am on [macOS Apple Silicon / macOS Intel / Windows / Linux, plus roughly how much RAM you have], and my comfort level with the terminal is [beginner / okay / confident].
Read the repo's README and CONTRIBUTING files first, then tell me honestly whether I should use the prebuilt installer or build from source for my setup. If a prebuilt build exists for my machine, say so and stop rather than making me compile something unnecessarily.
If building from source is the right call, walk me through it one step at a time. Check which prerequisites I already have before installing anything new, explain in plain language what each command does before you run it, and pause after each step so I can confirm it worked. If a command fails, read the actual error, tell me what it means, and fix it rather than starting over.
Once it is running, this is the prompt that makes your reference clip actually sound like you.
🎙 write my voice reference script
I am recording a 30 second reference clip to clone my voice in Voicebox. Write me a script to read aloud that is designed to capture the full range of my speech rather than one flat tone.
Cover a mix of sentence types: a couple of statements, a question, something with genuine warmth, and something read a little faster. Include a spread of common sounds and a few numbers, and keep the vocabulary conversational so I do not slip into a stiff reading voice. Mark where I should pause.
The recording is for [what you will use it for, for example narrating your own videos], so match the energy to that. Then give me three things to check in the playback that would tell me I need to re-record.
the good part
Give Claude a voice
Voicebox ships a built-in Model Context Protocol server, which means Claude Code can call it directly: speak out loud in your cloned voice, transcribe audio, and list your captures and profiles. This is the part almost nobody in the comments knows about.
▶️ Leave Voicebox running. The MCP server lives inside the app, so the app has to be open for Claude to reach it.
🔌 Connect it to Claude Code. Run this in your terminal: claude mcp add voicebox --transport http --url http://127.0.0.1:17493/mcp --header "X-Voicebox-Client-Id: claude-code"
🗣 Use it. You now have four tools available: voicebox.speak, voicebox.transcribe, voicebox.list_captures and voicebox.list_profiles. Ask Claude to read its summary out loud when a long task finishes and it will, in your voice.
My favourite use for this: a long build or research run finishes while you are away from the desk, and Claude tells you out loud what it did instead of you scrolling back through a wall of output.
read this bit
Two things to get right
Only clone a voice you have the right to clone. Yours, or someone who has actually said yes in writing. Cloning a person's voice without permission is a real legal and ethical problem, not a grey area, and it is the fastest way to turn a fun tool into a serious mess.
MIT covers the app, not every model. Voicebox itself is MIT licensed, so the software is genuinely free to use and modify. The speech models it downloads each carry their own separate licence, and some are more restrictive about commercial use than others. If you plan to put generated audio into anything you sell or run ads against, open the licence for the specific engine you used and check it first.