How voice cloning works

Ten seconds is enough, and why

A voice model has already heard thousands of speakers. It has learned what changes from one voice to the next — pitch, timbre, the shape of the vowels, the pace — and what does not. Your recording does not teach it to speak; it only tells the model where your voice sits among all the ones it knows.

In practice that takes three parts. An encoder listens to your clip and turns it into a short sequence of numbers that describes how you sound, not what you said. A text-to-speech model is conditioned on that sequence and on the text you type, and predicts the next sliver of audio, twelve and a half times a second. A decoder turns those predictions back into sound. The tool on this site runs Kyutai’s Pocket TTS, a model of 100 million parameters — small enough to do all three on a laptop’s processor, without a graphics card.

When it is legal

Your own voice: yes, everywhere we know of. Someone else’s: only with their explicit permission, and even then what you make with it is regulated. As of October 2026:

• United States, phone calls. Since February 2024 the FCC treats AI-generated voices as “artificial” under the Telephone Consumer Protection Act, so a call using one needs the called person’s prior consent (FCC announcement).
• Tennessee. The ELVIS Act, in force since 1 July 2024, makes your voice part of your protected likeness: using a simulation of it without permission is actionable, and so is offering a tool whose primary purpose is doing that (Vanderbilt Law School).
• European Union. Since 2 August 2026, AI Act Article 50 requires deepfake audio — generated speech that would pass as a real person — to be disclosed as artificial (Article 50 guide).
• Denmark has proposed giving every person a copyright-style right over their own voice and face, so that a clone made without consent can be taken down (Schjødt).

The model’s own licence says the same in one line: no cloning without explicit, lawful consent. That is why the tool asks you to confirm before it uses a recording. None of this is legal advice.

Cloning without uploading your voice

Most voice-cloning services work by uploading your recording and keeping the voice on their servers — that is the product. A voice is a biometric you hand over just by talking, so it is worth knowing that the alternative now runs in a browser tab.

Our voice cloning tool downloads the model once and learns your voice on your device, in English, French, German, Spanish, Portuguese or Italian. There is no account and no length limit, and your files are never uploaded — open the Network tab while it works and you can watch that for yourself.

Questions people actually ask