Clone once, then speak over keepers — not over every sample

A clone is a reusable id, not a video. clone_voice_from_audio takes a sample you own (or have consent to use) and returns a Cartesia voice_id. Later generate_speech and change_voice calls can use it. Cloning bills once per sample. The leak is the sequel: reading every candidate take in "your" voice, then throwing the takes away, plus cloning again from a better sample when the first clone was noisy.

Clone your voice for videos is the job. list_cloned_voices is how you see what you already have. The voice cloning tool is the catalog door. The id is not universal across every TTS model in the app — Cartesia-backed calls only.

Sample quality is the first lock

Background noise and extra speakers in the sample degrade the clone. Do not clone the scratch VO from take two because it is handy. Clone a clean room, one speaker, then stop. A second clone from a "better" scratch is a second bill and a second identity. Pick one.

Consent is not optional. Only clone voices you own or have explicit permission to use.

The clone is setup. It can happen before a single video exists. What cannot happen first is using it as wallpaper on a bin of unlocked pictures. Each generate_speech is billed. Each attach is billed. "It sounds like me" is not a reason to narrate a miss.

Use the id on files that will ship

After picture lock (or a duration you will not change):

1. generate_speech with the cloned voice_id.
2. attach_audio_to_video on the keeper.
3. Caption the result if the destination is silent.

change_voice is the cousin: same timing, different Cartesia voice, on an existing recording. That is also a later step, and it needs a performance you are keeping.

Do not clone as a way to preview a script. You can read the script. Pay speech when the cut holds it.

What this does not do

It does not lipsync a face. It does not dub a language. It does not make a noisy sample into a studio. Isolation (isolate_audio) can rescue a keeper VO before you clone from it — still on a sample you intend to be the source of identity, not on a throwaway.

estimate_cost for the clone and for the later reads. The clone is the cheap-looking half. The reads are the habit.

FAQ

Should I clone before I have any video?

You can, from a clean sample. Do not then feel obligated to generate speech for every experiment. The id waits. The pictures should too.

Can I use the same voice_id on every model?

No. Cartesia-backed generate_speech and change_voice. Other TTS rows will not treat it as a universal speaker.

Why not clone from the first decent take's audio?

Because that take may have room tone, a laugh, a second person. Clone from a dedicated sample. Use the take as a video, not as a voice source, unless it is actually clean.

Does listing clones cost a generate?

list_cloned_voices is a lookup. Generating speech with a listed id is the meter. Look up first so you do not clone a duplicate of a voice you already have.