Free voice cloning, in your browser.

At a glance

How zero-shot AI voice cloning works

Zero-shot cloning means the model has never been trained on your voice; it hears a sample once and imitates it. Chatterbox does this in three steps. A speech encoder listens to your reference clip and turns it into two things: a compact description of the voice (its pitch range, timbre and the way it sounds in its room), and the clip itself as a string of speech tokens. A language model then reads your text and writes new speech tokens, one for every 40 milliseconds of audio, continuing in the style of the reference. Finally a decoder turns those tokens back into sound in the reference voice.

All three run here, on your graphics card, through WebGPU. The model is downloaded from Hugging Face the first time, about 560 MB, and kept by the browser, so the second visit starts in seconds. Your recording is resampled to 24 kHz, trimmed of silence at the ends and cut to its first five seconds of speech; it is never sent anywhere, and closing the tab discards it.

How much audio do you need to clone a voice?

Five seconds of one person speaking clearly is enough here, and three is the minimum; a longer sample does not improve the likeness, for a reason explained below. What the reference sounds like matters far more than its length. Record in a quiet, soft room: a bedroom or a car beats a kitchen, because hard surfaces add echo, and the model copies the room along with the voice. Hold the phone or microphone a hand's width from your mouth and speak as you want the result to sound. If you read flatly, the clone reads flatly. Only the first five seconds of speech are used: the decoder re-reads the whole reference for every fraction of a second it renders, so a longer one makes every sentence slower without helping the likeness much. Record a little more than that and start speaking straight away. A clip with music, a second voice or heavy noise underneath gives a muddy clone; clean it first with noise reduction or pull the voice out of a song with the AI vocal remover.

Write the text the way it should be spoken, with punctuation, and keep sentences to a reasonable length; each one is generated separately and played as it arrives. Chatterbox Turbo also understands a few non-speech tags inline: [laugh], [chuckle], [cough] and [sigh].

Consent and honesty

A cloned voice can be used to deceive: fake phone calls to relatives, messages put in someone's mouth, attempts to get past a bank's voice check. Do not clone anyone without their permission. The page will not start until you confirm that the voice is yours or that you have permission, and every file it writes carries a tag saying it is a cloned, synthetic voice. That tag is metadata and easy to strip, so it is a label, not protection. The responsibility is yours. In many places, using someone's voice without consent is unlawful as well as wrong.

What it can't do

It needs WebGPU. On a phone, or in Safari or Firefox without it, the page says so instead of crawling for minutes on the processor. It speaks English only. The likeness is good, not perfect: people who know the voice well will usually hear the difference, especially in emotion and rhythm. For a neutral narrator voice with no cloning, the text to speech page is faster and needs no GPU. Commercial services such as ElevenLabs offer more languages and fine-tuned "professional" clones, but they work by uploading your voice to their servers.