Clone the voice, then run lipsync

Clone the voice first. Then lipsync. Cartesia Voice Clone is 8 credits, voice-clone: upload a short sample, get a reusable voice for TTS. Clone your voice for videos is the task door (clone_voice_from_audio → voice_id → generate_speech). Then drive a mouth: Kling Avatar Pro is 58 credits, 4K, audio-to / image-to, still required; VEED Fabric 1.0 is 40 credits, 720p, same input shape. Consent is the gate. A celebrity sample you do not own is not an 8-credit problem.

This page is clone-then-talker. It is not video-to mouth-pass on an existing clip, and it is not dub-without-a-new-mouth.

Clone only a voice you have the right to use

The card: Cartesia, content type audio, category voice-clone, 8 credits. Description: clone any voice from a short sample for TTS. “Any” is a product sentence, not a licence. Only clone voices you own or have explicit consent to use. Founder, staff creator, a talent agreement that names synthetic use. Not a podcast rip. Not a dead relative without the authority that jurisdiction requires. Not a competitor’s ad.

clone_voice_from_audio:

1. Clean sample — one speaker, little bed, representative read. Noise and two mouths degrade the clone.
2. name required; description and language optional so you can find it later.
3. Receive voice_id. Reuse with Cartesia-backed generate_speech or change_voice. It is not a universal id across every TTS model in the catalog.
4. list_cloned_voices to browse instead of re-cloning every video. Cloning is billed per sample submitted. TTS after that is the speech tool’s normal rate. Confirm with estimate_cost.

Cartesia Sonic 3.5 is 5 credits TTS if you wanted a stock voice instead of a clone. Gemini 3.1 Flash TTS is 4 credits with inline audio tags. Clone is for identity. Stock is for a voice that does not have to be a specific person.

N.Y. GBL § 396-b (in force 9 June 2026) covers synthetic performers that appear as a human not recognizable as an identifiable natural person. A clone of an identifiable person is a different legal problem — publicity, agreement, sometimes a digital-replica statute — not that NY ad-disclosure object. Do not treat a $1,000 / $5,000 NY penalty table as the whole risk for a founder clone. Get the release. Disclose when the destination or the jurisdiction requires it (EU Article 50 from 2 August 2026 for AI that would pass as real at first exposure; YouTube photoreal overlay, 27 May 2026, on Shorts).

If the sample is the founder and the still is the founder, you still need both consents: voice and face. A still from a photoshoot does not automatically include a synthetic voice clause.

TTS the line, then pick the lipsync row from the files

Write the script. Generate speech with the voice_id. Time it out loud. Duration of the wav is duration of the mouth.

Then inventory the picture:

• Still + this wav → image-to / audio-to lipsync. Kling Avatar Pro at 58 credits, 4K, when delivery is 4K or a close-up. VEED Fabric 1.0 at 40 credits, 720p, when a feed file is enough. Fabric 1.0 Text at 40 credits, 720p, if you have a still and a script and you will accept auto voice — skip that row if the whole point was the clone. You already have a wav. Use Fabric 1.0 (audio-to), not Text.
• Existing talking video + this wav → video-to. Kling Lipsync 7 credits 4K, VEED Lipsync 4 credits. You would not have needed Avatar Pro. This page’s title assumed you were building a talker. If the plate already talks, do not pay 58.

Lipsync is mouth to track. Native-audio T2V will not speak the clone. Veo will invent a cousin of the line in a cousin of the voice.

Procedure:

1. Consent on file.
2. Clone (8 credits).
3. generate_speech for this script. Listen. Re-take the TTS, not the clone, if the read is wrong. Re-clone only if the sample was the problem.
4. Lock the still (or confirm the video plate).
5. Run Avatar Pro or Fabric with still + wav.
6. Watch teeth and plosives. Caption after, from the wav you attached. Add captions to video.

Do not clone, then paste the script into a native-audio prompt “to save the lipsync.” You will lose the identity you paid 8 credits to capture.

Do not clone a voice and lipsync a different face as a fake testimonial. That is the trust problem UGC stats are about (Kapwing 2026: 92% trust peers) — and it is a deception problem besides.

8 + 58 is not 8 + 4

Budget honestly:

• Clone 8, then Avatar Pro 58, then captions. 4K talker from a still and a consented voice.
• Clone 8, then Fabric 40, then captions. 720p talker.
• Clone 8, then video-to 7 or 4, if a talking plate already exists.

TTS-only (clone 8 + Sonic 5 or Flash TTS 4) on a silent B-roll is cheaper when the mouth is not on screen. If you cannot see lips, skip lipsync. A pack orbit does not need visemes.

AI voice cloning is the clone door. The lipsync row is a second click. Mixing them in one prompt is how you skip consent and skip the still.

Consent, clone, speak, mouth

The test: would the person on the sample sign a one-line email that you will clone this take and drive a still of them saying future scripts? If no, stop. If yes, keep the email with the voice_id.

ElevenLabs Voice Change is 5 credits, audio-to-audio — a different row if you already have a performance and want a different voice on it. This page’s order is clone → TTS the script → lipsync a still. Do not change_voice a celebrity interview and then Avatar Pro the founder’s face. That is two identities stacked. The still and the voice_id should be the same person, with the same consent file.

If you need a seat to clone a consented sample and run Avatar Pro or Fabric, sign up. Bring the wav and the still. Bring the yes.

FAQ

Does the clone work with every TTS model?

No. The voice_id is for Cartesia-backed generate_speech and change_voice. It is not a universal id on Gemini TTS or every other speech row.

Kling Avatar Pro or VEED Fabric after the clone?

Same inputs: still + audio. Avatar Pro is 58 credits at 4K. Fabric 1.0 is 40 credits at 720p. Pick the ceiling. Do not use Fabric Text if you already have the cloned wav.

Can I skip lipsync and generate Veo with the cloned voice?

Not as a catalog promise. Veo is native audio on its own performance. Attach the clone to a lipsync row (or to silent picture as a VO) if the identity of the voice is the point.