Cloning takes seconds of audio and no training run, which makes it easy to try and easy to misjudge. Almost every disappointing clone is a clip problem, and clip problems are cheap to fix — if you know what you are listening for.

​
The clip is the conditioning

This is the whole model in one sentence: every generation is conditioned on the reference clip you uploaded. There is no training step that averages away what is in it, and there is no separate “voice” extracted from it. Whatever the clip contains, the synthesized speech inherits. That explains every rule below without you having to memorize any of them:
  • Room reverb in the clip clones the room. The voice will sound like it is always in that room.
  • Background music clones a hum that follows the speaker into every sentence.
  • A second voice in the background is a second voice the model may reach for.
  • A clip recorded through a heavy codec clones the codec artifacts.
  • A shouted clip clones a shouted voice, and it will shout your terms of service.

​
What a good clip looks like

Longer is not better; clean is better. Ten seconds of clean speech beats two minutes with a door slam in it, because the door slam is now part of the voice. If you have a long recording, cut the best few seconds out of it rather than uploading the whole thing. Match the clip to the job. The clip sets register as well as timbre. If the voice will read support answers, record it reading something like a support answer — not a dramatic performance, and not a single word. Uploads are capped at 10 MB and must carry an audio/* content type. At these lengths that ceiling is generous; if you are near it, you are uploading more audio than helps.

​
Always send ref_text

When you don’t supply a transcript, the platform transcribes the clip for you and conditions on the result. A misheard word becomes part of the voice’s conditioning, and a clip that transcribes to nothing fails the clone outright with status: "failed".
This is the failure mode that most looks like a broken model and is actually a missing field. A noisy clip transcribes badly; the bad transcript is what the model is told the audio says; the clone comes out subtly wrong in a way no amount of retrying will fix, because retrying re-derives the same bad transcript. Send the exact words spoken in the clip and the problem disappears:
Transcribe it by ear, not with a tool. The point is to be right about a sentence you already know.

​
Judge it with preview, not with your own text

POST /v1/voices/{voice_id}/preview synthesizes one fixed sentence. You don’t get to choose the text, and that is the feature: two voices previewed on the same sentence are comparable, where two voices reading their own best lines are not. Listen for the things that come from the clip rather than from the model:
  • Does it sound like it’s in a room you didn’t intend?
  • Does the energy match what you want it saying at minute nine of a call?
  • Does it hold across a long sentence, or thin out at the end?

​
If it sounds wrong, re-record

Cloning is deterministic given the clip. Uploading the same audio again produces the same voice, so a retry is only worth it if you changed the input — a different take, a tighter cut, the transcript supplied, the noise gone. The fastest loop is: cut three candidate clips from one recording session, clone all three, preview them side by side, keep one and delete the rest. Deleting a voice erases the reference clip from storage before the row, so a rejected candidate leaves no voiceprint behind. A reference clip is a recording of a real person, which in several jurisdictions makes it biometric data and a protected likeness. consent_attested=true is required and the timestamp is stored on the voice — but an attestation records your claim, it does not verify it. Nothing checks who is speaking on the clip. Two consequences worth planning around:
  • Keep your own evidence of the speaker’s permission. It has to outlive the clip.
  • Voice clones expire three years after their last use, and that period is not configurable on any deployment. Using a voice resets the clock, so a voice in production never expires out from under you — but a voice you cloned once and shelved will go, along with its reference audio.
If you don’t need a specific person, use a built-in voice instead. They are designed from a text vocabulary rather than recorded from anyone, so they carry no likeness rights and no retention clock. See Compliance for the full picture.

​
Going further

Voices

The API, the reference-audio invariant, and voice design.

Clone a voice

End-to-end: upload, poll for ready, synthesize.