How it works
A voice-cloning model does not splice together samples. Out of the recording it extracts a voice identity — a string of numbers describing what makes that voice sound the way it does: the timbre, the speaking rate, the characteristics of the intonation. From then on it can say any text at all with that identity.
That is why so little of it is enough. The model already knows the general rules of speech; from the recording it only has to learn how this voice differs from the average. And that can be read out of a few dozen seconds of clean speech — which is the most uncomfortable property of the technology, because that much audio of almost anybody exists in public.
The two levels on offer
- Instant cloning. Half a minute to a minute of audio, immediate result. Recognisable, but the fine characteristics — the laugh, the rhythm of the pauses — do not come through.
- Professional cloning. Half an hour or more of studio-clean recording, and hours of processing. This one is similar enough to deceive.
- Verified cloning. At the professional level, providers typically ask for identity verification and a spoken consent statement: you have to read out a given sentence that records that you know about this and agree to it.
- Stock voice libraries. Not cloning: voices from the provider's own contracted speakers. Legally this is the cleanest route, and it is enough for most jobs.
Where the law actually bites
People reach for an analogy — "a voice is like a likeness" — and then, more often than not, they reach for copyright. That is not where this is decided. In the United States there is no single federal rule: the bill meant to settle the question was still in committee in mid-2026, so it is not in force. The states ran ahead. One state law has been in force since 2024, and its peculiarity is that it bans the tool itself, not only the use of the output; another state has regulated the contractual side since the start of 2025. The basis in all of them is the right of publicity — control over your own name, likeness and voice.
There is a second common mistake, on the data protection side. A voice is personal data, that much is true. But it is not automatically "special category" biometric data: it becomes that only when the purpose of the processing is specifically unique identification. A voice cloned for narration is therefore personal data but does not fall into the strictest bracket — voice-based access control does. Commercially the difference between the two is large.
What consent actually turns on
If you use the voice of a colleague or a client, consent is not a signature but the answer to four concrete questions.
- For what. Internal training? Advertising? Customer service? The phrase "for marketing purposes" is too broad to be worth anything.
- For how long. A fixed term, or until withdrawn. For the latter, write down how it can be withdrawn and what happens to the material already produced.
- Who has access. A cloned voice lives in an account. Who can generate with it, and how would you find out if somebody misused it?
- What happens after they leave. This is the most commonly forgotten point. A former colleague's voice can stay in the training material for years.
What the EU regulation requires
The relevant article of the AI Act has applied since 2 August 2026 — it is live. It contains two obligations that are independent of each other. One is the provider's: generated audio has to be marked in a machine-readable way. The other is the publisher's: if the content is synthetic audio or image capable of deceiving, that has to be disclosed to the public.
The sanction is not symbolic: the regulation allows fines of up to fifteen million euros or three percent of worldwide turnover, and each member state adds its own national enforcement rules and figures on top. For the practical details — when, where and how to label — see Labelling AI Video, the post that goes through it.
The European picture, which is instructive
In Europe the starting point is different, and a good deal clearer, because several national personality-rights laws name the voice outright instead of arriving at it by analogy. The Spanish organic law on the protection of honour, privacy and one's own image lists the name, the voice and the image together, and treats their use for advertising or commercial purposes as an unlawful intrusion. Hungarian civil law protects a voice recording on the same footing as a likeness: both making it and using it need the consent of the person concerned. There is nothing to argue about here — no consent, no use.
The most instructive item, though, is a 2025 US court decision that said two things at once: a voice as such is not protected by copyright, but a personality-rights claim stands. Seen from Europe that is familiar, because on this side the basis was never copyright either. Which means that anybody trying to fight the cloning of their own voice with a "copyright" argument is knocking on the wrong door.
What to do in practice
- If you do not need one specific person, do not clone. The providers' stock voice libraries are legally clean and good enough for most jobs.
- If you clone your own voice, write it down. Several platforms do not require a label when your own voice is used for dubbing — but this varies by platform, and it changes.
- Written consent, answering four questions. For what, for how long, who has access, what happens after they leave.
- Never clone from a public recording. A podcast, an interview or a social video is not consent — neither legally nor morally.
Common misconceptions
- "My voice is not protected, because it is not a work of art." The protection is not copyright but personality or publicity rights. That is a different body of law, and it does cover the voice.
- "A public recording is fair game, anyone can hear it." Being public is not consent. This is the most common and the most expensive mistake of them all.
- "If I label it as AI, it is fine." Labelling and consent are two separate obligations. A label does not replace permission.
- "A voice is always special-category biometric data." Only when the purpose is unique identification. For narration it is not.
What to take away
On the technology side the question is settled: half a minute of recording produces a recognisable clone. The decision is therefore not technical but legal and ethical. If you do not need one particular person's voice, use a stock voice. If you do, get written, specific consent — and never work from a public recording. On the workflow of multilingual dubbing, see Lipsync and Dubbing, which lays out the steps. And if your question is what to do when someone asks for money in your voice, the callback rule is the answer.