Faceless Desk
AI voice

Whose voice are you allowed to clone?

The tools clone any voice you feed them and never check; permission sits in a contract clause requiring the speaker's consent, which most never read.

Research-led · from vendor docs, not hands-on use 8 min read Updated 2026-09-10 Contains affiliate links

Disclosure: some links on this page earn us a commission if you subscribe, at no extra cost to you. It never changes what we recommend — where a free or cheaper tool wins, we say so.

Voice cloning is the one faceless-video capability where the software will happily let you do something you have no right to do, and tell you nothing about it. Upload thirty seconds of audio, get a usable clone. No check, no prompt, no consent step.

That silence is not permission. It is the vendor moving the problem to you, deliberately and in writing.

What the contract actually says

From ElevenLabs’ terms of use, the clause that governs this:

“a voice you are authorized to share with us”

That is the standard for creating a voice model — authorized, not available, not findable on YouTube. And the obligation is placed squarely on the user:

“You may not provide Input or create Output for which you do not have all the rights necessary to grant us the license described above.”

You also warrant that your content and voice models “will not violate any rights of any person or entity”, and you indemnify the vendor if that turns out to be false.

Read those together and the arrangement is clear. The tool does not verify consent because the contract makes consent your problem. If you clone a voice you had no right to, you have not defeated a safeguard — you have breached the agreement and accepted liability for whatever follows.

The second vendor says it in writing, literally

It is worth checking whether this is one vendor’s caution or the industry’s position, so here is Murf AImurf.ai · affiliate, whose terms of service are if anything more explicit. Its cloning clause is built around a defined term, the “Consenting Speaker”: a cloned voice is “an AI model trained on voice recordings of a Consenting Speaker”, and by uploading one you “represent and warrant that you have obtained the explicit written consent of any third-party Consenting Speakers”. Then the flat sentence: “The submission of a third party’s unauthorized voice recordings as training audio is prohibited.”

Explicit written consent is a higher bar than ElevenLabs’ authorized. It is also the bar the commissioned-voice case below turns on, so it is useful to see a vendor name it.

Two smaller things from the same reading. Murf’s marketing page answers “Can I clone a celebrity voice?” with “Yes”, followed immediately by a warning that doing so without permission “can lead to intellectual property infringement or even identity theft”; the contract, not the FAQ, is what you agreed to, and the contract says no. And on Murf, custom voice clones are listed as an Enterprise add-on rather than a feature of the Creator and Business plans, so for most channels the question is moot there until it is not.

Clauses read from murf.ai/legal/terms-of-service (effective 2024-12-19) and murf.ai/voice-cloning on 2026-09-08.

Two vendors using the same defined term is a pattern; three is the industry’s settled position. Descriptdescript.com · affiliate’s terms of service reach for the identical phrase — a “Consenting Speaker” is a person who “has given both you and Descript permission to use recordings of their voice to train an AI Voice”, and an “AI Voice” is a model trained on that person’s recordings. Word for word, the same construction as Murf’s.

But Descript adds something neither of the other two has, and it is the most interesting clause of the three. Consent is not only a warranty you assert on a form. §9.3 says you create a voice “by having a Consenting Speaker read a designated consent statement” — a defined term, capitalised, the “Consent Statement”. The permission is captured in the training audio itself: the speaker’s own voice, on the recording, agreeing to be cloned.

That is a materially better design than a checkbox, and worth saying so on a site that mostly finds these clauses tilted the vendor’s way. ElevenLabs and Murf ask you to promise you had permission. Descript asks the person whose voice it is. A Consent Statement is evidence that survives the argument; a ticked box on your account is not.

Two caveats keep it from being a solved problem. The terms allow the other route too — “or in some cases by submitting voice recordings of a Consenting Speaker” as training audio — which is the ordinary upload-a-file path, with the ordinary warranty behind it and no recorded consent in the file. And the same section grants Descript broad use of what you upload: to train the voice, to hold in de-identified research datasets, and to let employees, vendors and contractors listen to samples “to test the quality of your AI Voices and monitor for potential misuse”. Reasonable purposes, plainly stated, but a third party you did not name will hear the recording, which is itself something the speaker should know before they read the statement.

For a faceless channel the practical upshot is unchanged and now has a mechanism attached: if you are cloning someone else’s voice, get them to record the Consent Statement rather than tick a box on their behalf. If the tool you are on does not offer one, write the equivalent yourself and keep it.

Clauses read from descript.com/terms (last updated 2025-02-25), section 9.3, on 2026-09-10.

The four cases, in order of how safe they are

Your own voice. Unambiguous, and worth saying out loud: cloning yourself is the one case with no rights problem at all. For a faceless channel this is often overlooked, because “faceless” gets conflated with “not my voice”. It does not have to. A clone of your own voice keeps the format and removes the entire licensing question.

A stock voice from the vendor’s own library. Also clean. The vendor licensed those voices and offers them under your subscription — that is what you are paying for. Check whether the licence covers commercial use on your plan, which is a separate question with its own trap on free tiers.

A voice you commissioned. Fine, if you got it in writing, and the writing has to say more than “you may use this recording”. Cloning is not the same permission as narration. A voice actor who recorded fifty lines for you has not thereby agreed to a synthetic model of their voice producing unlimited lines forever. Get consent to create and use a voice model, name the scope, and keep the document.

A voice you found. No. A podcast host, a YouTuber, an actor, a public figure, someone in a clip — none of these have authorized anything. This is the case people actually ask about, and it fails at the first clause.

“But it is a public figure / parody / only for practice”

Three things are worth separating, because the answer differs.

The contract. Fails immediately in all three cases. You were not authorized, so you breached the terms regardless of your purpose. That is true even if nobody ever hears the output.

Publicity and likeness rights. A voice can be protected as part of a person’s identity in many jurisdictions, and several have recently strengthened exactly this protection in response to AI cloning. Being famous does not weaken this — it usually makes it more valuable and more actively defended.

Platform policy. YouTube requires disclosure of realistic synthetic media, and separately prohibits content that impersonates a person. Disclosing an unauthorized clone does not fix an unauthorized clone; it just documents it.

Parody has genuine legal protection in some places, in some forms, in ways that do not travel across borders reliably. It is not a general-purpose exemption, and a faceless channel monetizing at scale is a poor test case for one.

What this means in practice

The uncomfortable version: most of the risk here is invisible until it is not. Cloning without permission produces no error message, no strike, no immediate consequence. The consequence arrives later, aimed at a channel that by then has a back catalogue and an income.

So the practical position is not “avoid cloning”. It is:

  • Clone yourself, or use a licensed stock voice. Both cover the overwhelming majority of faceless production, and both are already paid for.
  • If you commission a voice, get cloning consent explicitly and in writing. Scope it: what it can be used for, for how long, and whether it can be revoked.
  • Keep the paperwork with the channel, not in an inbox. Whoever needs it later — a platform, a buyer, a lawyer — will want it years from now.
  • Do not clone a voice because the tool let you. The tool lets you because the contract already moved the liability.

What we cannot tell you

Whether a specific use is lawful where you live. Voice and likeness protection varies enormously by jurisdiction, is changing quickly in response to exactly this technology, and turns on facts about your use that no article can know.

What is stated above is what the vendor’s own terms say and what the platform’s own policy says. Both are public, both are checkable, and both are more restrictive than most people assume before they read them. If a project depends on the answer, read your own vendor’s current terms — they change — and take actual advice.

Caveat: this is a reading of published terms and policy, not legal advice. Terms change without notice; the ElevenLabs clauses were read on 2026-09-01, Murf’s on 2026-09-08 and Descript’s on 2026-09-10.

Sources