Consent-first voice cloning, explained

Your voice is about the most personal thing you can hand to a product. A voice twin that writes and speaks like you is genuinely useful — and it is exactly the kind of feature that goes wrong when a company treats consent as a checkbox in a PDF nobody reads. So before we shipped voice twins, we decided what the gate would be, and we put it in the product itself. This essay explains that gate, end to end, including the one part that is not finished yet.

The gate lives in the product, not the paperwork

The paperwork is real: our biometric and likeness data consent policy spells out what we collect, why, how long we keep it, and how it is destroyed. But policies describe behavior; they don't enforce it. The enforcement is in the recording flow.

When you train a voice twin, the recording screen asks for a specific, separate consent — separate from accepting our terms, and scoped only to biometric processing. It is not pre-ticked, and it is not optional. If that confirmation is missing, the backend refuses the request outright and tells you why: "Voice cloning requires your explicit consent. Please confirm you are the owner of this voice and consent to creating a synthetic model." Nothing is uploaded to a cloning pipeline, nothing is trained.

When you do consent, that fact is stored with a timestamp, attached to the exact voice version it covers. Consent here is not a mood — it is a record with a date on it.

You say the consent out loud

Ticking a box is easy. So the pipeline also expects something harder to fake: a consent recording, in the same voice being cloned, speaking the consent statement word for word:

"I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model."

The wording is fixed on purpose. The synthesis engine behind voice twins — Google Cloud's custom-voice API, listed openly on our sub-processors page — rejects paraphrases and custom scripts. That rigidity is a feature: the consent artifact is a person, in their own voice, naming exactly what they are agreeing to. It travels with the sample, and it is part of what gets destroyed when you leave.

Own voice only — even with permission

This is the rule that costs us the most, and the one we are keeping: you may not build a twin of another person on your account. Not your co-founder, not a client who emailed you an enthusiastic yes. Own voice, own likeness, adults only — it is in the consent step, in the policy, and in our acceptable use policy.

Why refuse even permissioned cloning? Because "I have their permission" is unverifiable at the moment it matters, and the failure mode — a synthetic voice of someone who never agreed — is the single worst thing this category of product can produce. A rule with no exceptions is a rule the product can actually enforce.

Where this stands for visual twins

Honesty requires a status report, not just principles. For voice twins, everything above is live in the product today. For visual twins, the equivalent in-product consent step is still being rolled out. Until it ships, uploading reference photos is your confirmation that the face is your own and that you agree to the policy — and you can have that data destroyed at any time by emailing privacy@repic.site.

We would rather publish exactly where the gate stands than imply one that is not there yet. When the visual consent step ships, this essay gets updated, and the policy's status callout goes with it.

Deleting is as real as creating

Consent you cannot withdraw is not consent. You can delete a voice version from the product, and withdrawing consent stops any further use and triggers destruction of the underlying sample, the consent recording, and the derived voiceprint, on the schedule set out in the policy. Withdrawal does not undo processing that already happened — nothing can — but it ends it.

That is the whole design: an explicit consent step the product enforces, a spoken consent artifact in your own voice, a no-exceptions own-voice rule, an honest status line for the part still rolling out, and deletion that actually deletes. Consent-first is not our compliance posture. It is the product decision that makes a personal-brand tool trustworthy enough to exist.