Consent and Disclosure When Using a Cloned Voice

Someone recorded a narration for you last year and was happy with the result. This year you need three more lines, they are unavailable, and you have sixty minutes of their voice sitting in a project folder. The technical question — can a model be built from that — has an easy answer. The one that matters is whether they agreed to it, and almost certainly they did not, because nobody asked.

This is the ordinary shape of the problem. It is rarely deliberate misuse; it is a permission that was granted for one thing being quietly extended to another. This guide covers what a consent for synthesis has to cover, how to think about disclosure, and a policy you can adopt so that the decision is made before the deadline rather than during it.

Recording Consent Is Not Synthesis Consent

A standard release covers using a recording. Synthesis is a different act: it produces new speech that the person never said, in their voice, indefinitely, in contexts that did not exist when they agreed.

The distinction is easiest to see by asking what changes. A recording is finite — it contains what it contains. A model is generative, and the person has effectively agreed to statements not yet written. Most people, asked plainly, understand that difference immediately and have views about it.

So the permission has to be asked for separately and in those terms. Not "may we record you" but "may we build a voice model from this recording, and use it to generate speech you have not said".

What a Usable Consent Covers

The purpose of writing it down is not formality. It is that in eighteen months nobody will remember what was agreed, and the written scope is what everyone can return to.

Two of these are routinely skipped and cause most of the later difficulty. Duration, because an unlimited grant seems simpler at the time. And withdrawal, because nobody wants to discuss the end of a relationship at the start of one.

There is also a practical point about who is asked. Consent has to come from the person whose voice it is, not from whoever holds the recording or commissioned it. Owning a file is not the same as having permission to build a model of the speaker in it.

Disclosure

The question of whether to tell the audience is separate from consent, and it is decided by a different test.

Requirements vary by platform and jurisdiction and are actively changing, so the current rules where you publish are something to check rather than to remember. The durable test is simpler: would a reasonable member of your audience feel misled if they found out afterwards? If the answer is yes, disclose, whatever the rules currently say.

That test usually produces clear answers. A synthetic narrator on a technical explainer is not something most viewers would object to. The same voice giving a personal testimonial, a recommendation, or an apparently spontaneous statement is a different matter, because the audience is being invited to trust a person.

Where disclosure is warranted, place it where it will actually be seen — in the video itself or at the top of the description, not in a footer. And use plain language. A line saying the narration was generated using a synthetic version of a named person's voice, with their permission, tells the audience everything relevant in one sentence. The general principles for labelling generated media apply here too, and are covered in disclosing AI-generated media to your audience.

Cloning Your Own Voice

Using your own voice removes the consent question and leaves the disclosure one intact.

An audience that believes they are listening to you speak is forming an impression about the recording. In informational content that impression rarely matters. In anything personal — a direct address, a statement of opinion, a message to customers — it does, and the audience's reaction on learning the truth is the thing to predict.

There is a second consideration that catches organisations rather than individuals. A model of your voice held on a company's systems outlives your involvement with that company unless something says otherwise. If you are cloning yourself in a work context, the scope questions above apply to you as much as to anyone else, and you are the person who benefits from having answered them.

The Cases That Are Genuinely Hard

Most of this is straightforward once the questions are asked. Three situations are not, and they are worth thinking about before they arrive.

Someone who has left. A model built with valid permission, and the person is no longer connected to the organisation. Whether the permission continues depends entirely on what it said about duration, which is why the term clause matters more than it seems at signing. Without one, you are relying on an interpretation nobody agreed to.

Content the person would not have said. The voice is technically permitted for this channel, and the specific script is something the speaker would object to. A scope defined by channel rather than by content type does not catch this, which is an argument for defining both. The honest test is whether you would be comfortable showing them the script — if not, the permission is not really covering it.

Archive material. Recordings of someone who cannot be asked, either because they are unreachable or because they have died. There is no procedural answer here. What there is, is a question about whether the audience would understand what they were hearing, and whether the people close to that person would recognise the use as respectful. Both deserve more weight than the technical possibility does.

None of these have clean answers, and the point of writing a policy is not to solve them in advance. It is to make sure they are noticed as decisions rather than made by default because a model already existed and a deadline was close.

Where the Recording Itself Comes In

One practical note that connects to the technical side. A consent obtained after a recording exists is harder to scope, because the recording was made for something else and the person is being asked to extend it.

Getting this right is easier when the reference session is planned as a reference session. The speaker knows what it is for, the scope conversation happens before anything is recorded, and the material can be structured for the purpose rather than salvaged from a narration. What that session should actually contain is covered in recording a clean reference clip for voice work.

A Policy Worth Adopting

The point of writing this down in advance is that the alternative is deciding under pressure, on a deadline, with a project half-finished.

None of this is onerous once it exists. The cost is a single afternoon and a template; the cost of not having it is a conversation you will have at the worst possible moment.

Next Steps

Check what permission you actually hold for any voice recordings you are already using. In most organisations the honest answer is a general release that never mentioned synthesis, which means the question has not been asked rather than answered.

Then write the consent template and the disclosure line before your next project needs them. Both are short documents, and having them turns an uncomfortable negotiation into a routine one. The technical side — what a clone can and cannot do, and what the reference recording decides — is in voice cloning: what it needs and what it can't do, with the rest under AI voice and TTS.