Clone a voice with AI, then write what it says next.

Your own, or one you have explicit permission to use. Cloning a voice is the one capability here where the technology is not the hard part — the permission is. A public figure’s voice, a colleague’s recording off a call, a sample lifted from a podcast: all technically possible, none of them yours to use. Get it in writing before you clone, not after.

The page was built from a single eight-second clip, which is enough to be recognisable. Longer and cleaner is better — one speaker, no music, no crosstalk — and if the only recording you have is noisy, isolating the speech first is a separate step you can run on the same key.

Yes, with the multilingual models — the same cloned voice id, a language on the request, and a different script. Around forty languages between the ElevenLabs and MiniMax tiers. What it cannot do is give you an accent you never had: a cloned English speaker reading Spanish sounds like an English speaker reading Spanish, which is either exactly what you want or exactly what you do not.

Within a maker, yes — one clone spoke through two different ElevenLabs models on this page with no re-cloning. Across makers, no: an ElevenLabs voice cannot be handed to a MiniMax model, and the API refuses it with that reason rather than degrading quietly. Worth knowing before you standardise on a voice you may want to move later.

Pass the speech to an avatar model along with a photograph, and the mouth is driven by the audio. Both clips further up this page were made that way — one still from the original recording, two different pieces of cloned speech. Up to ten minutes in a single request.

Creating the clone is not itself a billed generation; speaking with it is, by the character, at a rate that depends on which speech model you choose. Putting it on a face is billed by the second of video. Each model page carries its own rate.