Voice cloning and consent: what creators should agree to

This article is general information, not legal advice. Voice, privacy and contract rules differ by country and change often. Get independent advice before signing anything.

Of everything that makes you recognisable, your voice is the cheapest to copy and the easiest to license carelessly. A face needs a camera, a studio and a shoot. A voice needs a few minutes of clean audio — and most creators have already published hours of it.

It is also the asset people most often hand over by accident, folded into an agreement that says “likeness” and never says “voice”. This article is about treating voice as its own deal: what a clone actually is, how the law is starting to handle it, and the terms to settle before you record a single line. It pairs with our general fair likeness licence checklist, which covers the clauses common to all likeness deals.

Why voice needs a separate agreement

Three things make voice different from appearance.

A voice clone can say anything, instantly. A video replica usually needs a script, a render and someone to approve the result. Synthetic speech can be generated in seconds, in bulk, from text typed by anyone with access. The gap between “licensed” and “unapproved” is one careless prompt wide.

Audio travels without its label. A video can carry a visible on-screen AI disclosure. A phone message, a radio spot, a podcast ad read or an in-app prompt has no screen to put it on. Your voice can end up somewhere unrecognisable as an ad at all.

Voice is used for identity. Nobody rings your bank with your face. Voice cloning sits in the middle of a live fraud problem, which is why regulators reached it first.

How the law is treating voice

Voice is rarely an afterthought in the newer rules — it is usually named.

Washington’s personality rights statute reaches anyone who uses a living or deceased individual’s “name, voice, signature, photograph, forged digital likeness, or likeness” in advertising without the rights owner’s consent, and its definition of a forged digital likeness expressly includes “an audio recording which is either persistent or transmitted in real-time of an actual and identifiable individual’s voice” (RCW 63.60.050; RCW 63.60.020). The US federal bill now moving through Congress would create a right in “the voice or visual likeness of an individual” (S. 4591) — see digital replica laws in the US for where that sits.

On phone calls the position is already settled in the US: the FCC ruled that “calls made with AI-generated voices are ‘artificial’ under the Telephone Consumer Protection Act”, putting voice cloning in robocalls under the same consent rules as prerecorded messages (FCC, 8 February 2024).

In the EU, synthetic audio is inside the AI Act’s transparency rules. Providers of AI systems “generating synthetic audio, image, video or text content” must mark outputs machine-readably, and deployers of a deep fake — which the definition covers in audio as well as video — must disclose it (Article 50). Our Article 50 guide has the detail.

Platforms are less consistent. YouTube asks creators to disclose content that “makes a real person appear to say or do something they didn’t do”, but explicitly exempts “cloning one’s own voice to create voice overs or dubs” from disclosure (YouTube Help). Useful to know: a brand using your licensed voice is not in that exemption.

Voice-cloning tools themselves often require proof. One widely used service’s professional cloning flow states “we only allow you to clone your own voice” and puts users through a recorded verification step before training (ElevenLabs docs). If a brand asks you to record samples for them to upload as their own voice, that is a warning sign about the whole arrangement.

Your recordings are data, and sometimes special data

The audio used to build your clone is a lasting asset that sits on someone else’s servers.

Whether it is biometric data depends on what is done with it. The UK regulator’s guidance is careful here: biometric data is personal data relating to physical, physiological or behavioural characteristics “which allow or confirm someone’s unique identification”, and it only becomes special category data if “you use it to uniquely identify someone” — so an audio recording on its own, without that technical processing, is not biometric data (ICO).

That distinction matters less than it sounds. Either way the recordings are personal data about you where data protection law applies, and either way the practical asks are the same: encrypted storage, a named list of who can access them, no use for training anything other than your own voice, and deletion of both the recordings and the trained model on request, with a deadline.

What to settle before you record

1. Voice as its own line item. Appearance, voice, name and signature should each be listed as included or excluded. If an agreement only says “likeness”, ask for voice to be written in or struck out.

2. Which vendor, and where the model lives. Name the tool used to train and run the clone, and say whether the model file can leave that vendor. A voice model held by three sub-processors is three times the problem.

3. What the clone may say. The single most important clause. Approve scripts, or an agreed claim set with categories you will never voice. Require fresh approval for anything resembling a personal endorsement — an AI version of you recommending a product you have never used is the exact risk that endorsement and testimonial rules exist to catch.

4. Languages and dubbing. Synthetic voice crosses languages easily. Decide whether your voice may be used in languages you do not speak, and whether you can review a translation you cannot check. Many creators say yes to one market and discover a campaign in five.

5. Singing, and anything musical. Treat singing as a separate grant. It carries different expectations, different rates and different rights.

6. Blocked categories. The list you would use for video applies to audio and should be attached to the agreement: politics, gambling, alcohol, weight loss, adult content, crypto and financial promotions, anything aimed at children, competitors of brands you already work with.

7. Territories, dates and no evergreen renewal. Name countries, set an end date, refuse perpetual or irrevocable grants. Short terms are the best protection you have.

8. Pay per licence and per renewal. One fee for an unlimited bank of generated audio is not a licence, it is a buyout. Ask for a record of every licence issued with your voice so you can check it against what you were paid.

9. Nothing is resold, nothing is retrained. Advertisers should receive finished, approved audio — never the voice model, never your raw recordings. No training of general models on your material, and no sub-licensing without your written permission.

10. Labels and provenance. Ask that output is labelled as AI where the format allows, and that provenance metadata travels with the file, so an audio clip separated from its ad can still be traced. See content credentials.

11. A misuse route that exists in advance. A named contact, a stated response time, the ability to stop new generations immediately, and support for takedowns — not a clause saying misuse is your problem.

Two things to do for yourself

Separately from any contract: if you license your voice, do not use voice as an authentication factor anywhere you can avoid it, and agree a private verification word with family. Voice cloning is actively used in impersonation fraud — the FTC ran a public challenge specifically to surface ways to prevent, detect and evaluate it (FTC), and its proposed extension of the impersonation rule to individuals is aimed at the same conduct (FTC).

And keep your own copy of the master recordings. If a relationship ends badly, the person holding the only clean audio has all the leverage.

The one-page checklist

  • Voice listed separately from face, name and signature
  • Vendor named; model may not leave it without my permission
  • Scripts or an approved claim set; endorsements need a separate yes each time
  • Languages and dubbing stated
  • Singing excluded, or priced as its own grant
  • Blocked categories attached to the agreement
  • Countries named, end date set, nothing perpetual
  • Fee per licence and per renewal, with a record I can check
  • No resale of the model, no raw files to advertisers, no other model training
  • Output labelled and provenance carried in the file
  • Named misuse contact and response time
  • Recordings and trained model deletable on request, with a deadline
  • I hold my own copy of the masters
  • Independent legal advice before signing

How Likeness approaches voice

Likeness is being built so voice is always a separate, explicit grant rather than a line inside a broader release: scope and blocked categories set by you, approval per brand, a payment for every licensed use, labelled output tied to a licence reference, and deletion of recordings and the trained model on request. It is in development and not yet open — see for creators and trust and consent, or email [email protected].


General information only, not legal advice. Laws and platform policies differ by country and change often; take qualified advice before signing a voice agreement.

Sources

Previous post
Next post