You Own the Recording. The Company Owns the Voice.

Ask who owns a voice clone and the argument circles, because three different objects wear the same name.

• You own the recordings, the vendor owns the voice clone built from them. Copyright in a recording you made sits with you by default, and consumer terms confirm as much: ElevenLabs states that as between you and ElevenLabs, you retain all rights in and to your Input.

• The trained model is the vendor's. Weights, a speaker embedding and reference data, derived by a machine from your audio plus the company's base model, held on infrastructure you cannot reach.

• The licence is the operative clause, not the ownership line. Most consumer grants are perpetual, irrevocable, worldwide, royalty-free and sublicensable through multiple tiers.

• Copyright does not protect a voice. Voice is protected mainly through right of publicity and biometric privacy statutes, because the sound of a person is not a copyrightable work.

• Acquisition and insolvency transfer both the data and the model. The assignment clause in the terms and the business transfer clause in the privacy policy are written already, and you agreed in advance.

• Deleting your account deletes files, not what a model learned. Machine unlearning is an open research problem, not a feature.

• Your own copy of the source audio is the one control that survives all of the above. Keep a copy, off the platform.

The generated audio is where your plan tier quietly decides things. Paid users typically get commercial use. Free users typically do not. And pure machine output has no human author, so it is not registrable for copyright in the United States: the D.C. Circuit affirmed the human-authorship requirement in Thaler v. Perlmutter in March 2025, and the Supreme Court declined to review the case in March 2026. You can sell synthetic speech in your own voice. You will struggle to stop anyone else copying the file.

So a company can tell you that you own your voice data and be telling the truth, while holding the only copy of the model, on servers you cannot reach, under a licence you cannot revoke.

The ownership line is the reassuring sentence, placed high. The licence grant sits further down, and the grant does the work.

Use the ElevenLabs grant, written in unusually plain language for the genre. The company takes a licence that is perpetual and irrevocable, nonexclusive, royalty-free and fully paid, worldwide, and sublicensable through multiple tiers. It permits the company to reproduce, modify, adapt, translate, distribute, publish, perform, display and build derivative works from your content, to provide the service, improve the service and develop new products.

Perpetual means no end date. Irrevocable means you cannot take the grant back, even after you close your account. Worldwide means no jurisdiction is carved out. Sublicensable through multiple tiers means the rights can be handed to a partner, who can hand them to their partner. Royalty-free means none of that movement generates a payment to you.

The concession comes first, because the criticism is cheap without one. Broad licences are not sinister. Cloud services genuinely need them, to cache, transcode, back up, replicate across regions and pass audio to subprocessors, and a narrow licence would break the product. I have signed grants like this as a customer and written scoped versions of them as a founder, and no working service takes nothing.

The trouble is that the wording permitting a backup also permits product development, and "improve our services" does enormous quiet work in most of these clauses.

Two words separate a fair grant from a greedy one, and both are searchable in thirty seconds. Purpose: service to you, or open-ended. Training: an explicit carve-out, or silence.

Almost nobody looks. Obar and Oeldorf-Hirsch put 543 participants in front of a fake social network called NameDrop: 74% skipped the privacy policy by taking the quick-join option, and 97% agreed to the policy anyway. Published in 2020, and we all recognise ourselves in the number. Voice is the one category where the habit costs something that cannot be reissued. A password gets rotated in a minute. Nobody has ever rotated a voice.

Ask a vendor to send you your voice model and you will get a polite no. Some terms even say the voice model is your intellectual property. You will still not get a file, for three reasons, and only one of them is about you.

The first is technical. A modern clone is often not a standalone artifact but a speaker embedding: a small conditioning vector or adapter that means nothing except against the vendor's proprietary base model. The embedding without the base model is a key to a door you do not have.

The second is commercial. The base model is the company's core intellectual property, and an export that reconstructs your voice elsewhere dissolves the moat.

The third is safety, and I will defend this one even though it costs my own users something they would like to have. A portable, downloadable clone of a named human being is a fraud kit. Vendors that gate cloning behind consent verification are not going to ship the artifact that defeats their own verification.

So the rule holds whichever logo is on the product. Model portability barely exists here. Your audio goes anywhere, and that portability is the whole of your leverage.

The law lands on the same side, in a slightly depressing way. GDPR Article 20 gives you the right to receive personal data "which he or she has provided to a controller." Data the controller derived or inferred, which is exactly what a trained model is, sits outside that right. Portability was written for the file you uploaded, never for the thing built out of your upload.

Voice is protected. Not by the mechanism most people reach for.

Copyright does not cover a voice. Copyright covers the recording, not the human instrument that made the sound. In Lehrman v. Lovo (S.D.N.Y., 10 July 2025), two voice actors sued an AI text-to-speech company after recordings they supplied through Fiverr, for what they understood was internal research, allegedly turned up as cloned voices in a commercial product. The court dismissed most of the federal copyright and trademark claims: vocal attributes were not protectable as trademarks where the voice functions as the product rather than as a badge of origin.

Right of publicity does cover a voice, including an imitation. Old, well-tested law with real money attached. In Midler v. Ford Motor Co. (9th Cir. 1988), Ford's agency hired a former backup singer to imitate Bette Midler after Midler turned the job down, and the Ninth Circuit held that a distinctive voice, deliberately imitated for commercial gain, is protectable. In Waits v. Frito-Lay (9th Cir. 1992), Tom Waits won on the same theory and the court affirmed an award of $2,375,000. Back in Lehrman, publicity is the branch that survived: claims under New York Civil Rights Law sections 50 and 51 went forward, alongside consumer protection and breach of contract claims.

Statutes are arriving, unevenly. Tennessee's ELVIS Act, signed 21 March 2024 and effective 1 July 2024, added voice to a state property right covering name, photograph and likeness, and reaches a simulation of a voice, not only the real thing, which made it the first state law written specifically for AI voice replicas. Not the first to protect voice at all: California's right of publicity has listed voice since a 1984 amendment to Civil Code section 3344. California's AB 2602 and AB 1836 took effect on 1 January 2025, and AB 2602 voids contract terms that let someone create and use a digital replica of your voice in place of work you would have performed, where the contract lacks a reasonably specific description of the uses and you were not represented by counsel or a union. Federally, the NO FAKES Act of 2026 (S.4591 and H.R.8915) was reported out of committee on 18 June 2026, and a reported bill is not a law. State reach is set out on the digital replica laws and how far each one goes, and the federal bill has what the NO FAKES Act would and would not change.

Biometric privacy is the sharper instrument. Illinois BIPA lists voiceprints as biometric identifiers and requires written notice and written consent before collection. In May 2026 nine coordinated class actions were filed in Illinois by broadcast journalists, voice actors, podcasters and audiobook narrators, one suit each against Adobe, Alphabet, Amazon, Apple, ElevenLabs, Meta, Microsoft, NVIDIA and Samsung. The complaints allege that voiceprints were extracted from publicly available recordings and used to train commercial AI voice models without the notice and written consent BIPA requires. Those are allegations, and none has been proven. The theory is being tested against nine of the largest companies in the industry at once.

Outside the United States, voice is often sensitive by default. Under GDPR, voice used for identification is biometric data inside the Article 9 special-category regime, which generally requires explicit consent. Under the Australian Privacy Act 1988, biometric information and biometric templates are sensitive information, and the OAIC's guidance is that consent is generally required before collection. Europe stacks labelling and disclosure duties on synthetic audio on top of that, covered on the EU AI Act's Article 50 transparency rules.

You do not hold your voice the way you hold a share certificate. You hold a right to object, after the fact.

Voice is the one asset where you keep the original, the vendor keeps the copy that matters, and the law only wakes up after somebody has impersonated you.

Start building your legacy Free build, 50 memories, no card required.

Consolidation in voice AI is routine, and users learn about a deal from a product update.

Descript acquired Lyrebird in September 2019, and Lyrebird's voice technology reappeared as Descript's Overdub feature rather than the standalone service people signed up for. Spotify announced its acquisition of Sonantic in June 2022, and the technology later surfaced inside Spotify's AI DJ. Veritone acquired VocaliD in June 2022. VocaliD had begun with personalised synthetic voices for people with severely impaired speech, blending whatever sounds a recipient could still make with recordings from a bank of volunteer voice donors. Those donors gave their voices to a mission. A mission does not automatically survive a change of control.

In none of those cases did the users choose the acquirer. They could not have. The terms had handled that already.

Two clauses decide the outcome, and both are already on your screen. The assignment clause in the terms of service, which typically says the company may assign its rights and obligations to a successor. And the business transfer clause in the privacy policy, which typically says your personal data may be transferred to the acquiring entity as part of a merger, acquisition, asset sale or restructure. ElevenLabs discloses that it may share personal data with "potential transactional partners" in connection with "the consideration, negotiation, or completion of a corporate transaction or re-organization."

Your consent to that transfer was collected in advance, before anybody on earth knew who the buyer would be. Not a loophole. The mechanism working as designed.

Insolvency is the harder case, and there is a live precedent over an asset more intimate than a voice. When 23andMe entered Chapter 11 in 2025, the asset on the table was a genetic database covering millions of people. Regeneron was declared the winning bidder in May 2025 at $256 million. Bidding reopened, Regeneron declined to raise, and on 14 July 2025 the TTAM Research Institute, founded by 23andMe co-founder Anne Wojcicki, completed the acquisition instead.

The instructive part is the regulator. On 31 March 2025, FTC Chair Andrew Ferguson wrote to the US Trustee stating that any purchaser should expressly agree to be bound by 23andMe's existing privacy policies, and that promises made to consumers must be kept through a bankruptcy sale. Call that the emerging norm for sensitive data in insolvency: the privacy promise travels with the asset. Then read the norm the other way. It had to be asserted, in one case, by one chair, in a letter.

Acquisition is the good outcome. The ordinary outcome is quieter. The API starts returning errors, support goes unanswered, the domain lapses, and the model is gone with nobody to ask.

Coqui, one of the more respected open speech companies, announced its shutdown in January 2024 and got the ending right almost by accident of principle. XTTS v2 had shipped openly in November 2023, two months before, and the code stayed on GitHub under the Mozilla Public License, so the models outlived the company and can still be self-hosted. Almost no shutdown looks like that.

If the only copy of your voice lives inside a running service, your voice has that service's lifespan. Most startups in this category will not outlast you. Mine included, on the actuarial evidence. The same clauses govern accounts you already hold and never think about in these terms: what happens to your ChatGPT memories when you die runs on the same assignment and business transfer language, with no export of the derived layer either.

You can delete your account. What disappears is narrower than the button suggests.

Deleting recordings deletes files. Deletion does not reach into a trained model and remove what the model learned from you. The research field trying to solve that is machine unlearning, and the problem is open. Surveys sort the work into exact approaches, which remove a data point's influence thoroughly by retraining, and approximate approaches, which merely minimise that influence at a fraction of the cost. The European Data Protection Supervisor tracks machine unlearning as an emerging technology to watch, which is the institutional way of saying nobody can reliably do this yet.

Regulators have named the gap. In Opinion 28/2024, published 17 December 2024, the EDPB set a high bar for treating a trained AI model as anonymous: only where the likelihood of extracting personal data from the model, directly or probabilistically or through queries, is insignificant for every data subject. The same Opinion lists, among measures a company might offer, erasure even where the Article 17(1) grounds do not apply, and a way to report that a model has regurgitated or memorised your data so unlearning techniques can be tested against the claim.

So ask for deletion in writing, in parts, and keep the reply. Are my source recordings deleted, backups included, and by what date? Is the trained model deleted at your synthesis provider, or only unlinked from my account, and who is that provider? Do any derivatives, aggregates or fine-tuned artifacts built from my audio survive? Can I download my original recordings, in what format? Will my audio train any model that serves other customers? If you are acquired or cease trading, what happens to my recordings and my model?

A vendor who answers all of that in plain sentences is worth trusting with your voice. Marketing language in reply is also an answer.

Retention windows repay a slow read too. ElevenLabs states that it will not keep data it generates about your voice longer than three years after your last interaction, and that it retains biometric data until the data is no longer needed or three years after termination of the relationship, whichever is sooner. That is a real, stated limit, which puts them ahead of every vendor that says nothing. Three years is also a long time to be somewhere you have stopped visiting.

Before you record anywhere, open the terms of service and the privacy policy and use the browser's find function. Nine searches, ten minutes, and you will know more than the homepage will ever tell you.

Search "retain" and "ownership". The document should say you keep rights in your input. If that sentence is missing, stop there.

Search "licence" and "license". Note whether the grant is perpetual, irrevocable and sublicensable, and whether it stops at providing the service to you or continues into developing new products.

Search "train". The sentence you want says your content will not train models unless you separately opt in. The sentence you do not want is an absence.

Search "voice model". Yours, theirs, or undescribed. Undescribed is the most common answer, and the answer means theirs.

Search "export" and "download". Model export will almost never be on offer. Audio export should be.

Search "delete" and "retention". Look for a stated window, named backups, and whether deletion covers the provider-side model or only your account record.

Search "merger", "acquisition" and "asset". Decide now whether an unnamed buyer inheriting your voice, under a policy they can amend, is a deal you accept.

Search "assign". The company can usually assign the agreement to a successor. You usually cannot assign anything.

Search "biometric". A vendor that names biometric law and states a retention period has thought about the problem. Silence on that word means nobody there has.

Then check your plan tier. Commercial use rights and free-tier restrictions are the most common surprise here, and they change most often.

That check tells you who you are dealing with. The next half hour protects you either way.

Record good source audio and keep the audio yourself. Thirty to sixty minutes of clean, varied speech, in a quiet room, on a decent microphone. Read something aloud, then put the page down and tell a story you have told a hundred times, because the reading gives you diction and the story gives you timing. Save uncompressed WAV where you can, on your own drive and in your own cloud storage, with a plain text file beside the audio saying who is speaking, when the recording was made, and what you consent to.

Do that and every vendor becomes replaceable, which is a stronger position than any clause will hand you. Cloning quality keeps improving, so good audio lets you rebuild on whatever the best system turns out to be in 2035. If the only copy lived inside a service that closed, there is nothing to rebuild from.

Two habits pair with that copy. One defensive, because good source audio is what an impersonator wants too: keeping your voice out of a cloning tool you never agreed to. One evidential, because a publicity claim starts as a feeling and has to become a filing: how to prove a piece of audio is really you. Undecided about recording at all? Whether voice cloning is safe to do is the better starting point.

Now run the nine searches on us.

Our terms say you retain ownership of all content, memories, media and data you put into the platform. The licence we take is limited, non-exclusive and royalty-free, scoped to processing, storing and using your content solely to provide the service to you. It does not extend to commercial use, or to model training except under a separate agreement you would have to sign. Voice recordings are handled as sensitive information, collected only with explicit consent, and you can withdraw that consent at any time, with the relevant data deleted within 30 days.

Three caveats, because a page like this one is worthless without them.

We do not build our own speech synthesis. Voice generation runs on a third-party provider, ElevenLabs today, behind an internal interface designed so the provider can be swapped. Your recordings live in our storage. The synthesis model does not. That provider is also one of the nine defendants in the Illinois voiceprint suits described above, which is a fact to weigh rather than one to bury: those allegations concern voiceprints said to have been taken from publicly available recordings, not customer audio supplied under consent, and none of them has been proven. We sit inside the same structural limit as everybody else here.

Our terms contain a business transfer carve-out, in the same place everybody else's does. If the company were acquired, your data could pass to the successor, bound by the privacy policy as written. Read that clause everywhere, including here.

And we are not offering you a model file, because we cannot. What we offer is control over release rather than control over weights. Executor Lock™ means Personas reach the people you nominate only after a seven-day, evidence-verified process you configured while alive, and Personas hold what you chose to put in, in your own voice, without inventing what you never said. Nobody in this category can promise you the weights, and anyone promising them is describing a product that does not exist.

So start with the recordings, because nobody can take those away. Fifty memories costs nothing, asks for no card at any point, and what you build never expires. For the paid tiers, compare what each plan includes. For the release mechanism, how Executor Lock releases Personas to the people you nominate sets out the process.

Voices are being modelled with permission and without. So the choice is narrower than the headline. Not whether a model of your voice will ever exist. Whether the version that outlives you was made on purpose, with consent, from audio you still hold a copy of.

Start building your legacy Free build, 50 memories, no card required.

Everything above comes from primary sources wherever a primary source exists: published vendor terms of service and privacy policies read in full, court opinions and appellate decisions, the text of state and federal bills, regulator documents from the EDPB, the FTC and the OAIC, and company announcements of the acquisitions described. Secondary reporting is used for dates, deal values and a small number of product details that no primary source records. What is not settled: the NO FAKES Act is a reported bill and not law, the 2026 BIPA class actions are unproven allegations, machine unlearning has no reliable technical answer yet, and vendor terms quoted here can be amended at any time without notice. Treat every date and status as accurate on the review date at the top of this page, and verify before relying on any of them.

If you find something on this page that is wrong or out of date, report the error through our contact page and we will correct the page and note the change.

• ElevenLabs Terms of Use and Privacy Policy, quoted for licence grant, biometric classification, retention periods and change-of-control disclosure

• Thaler v. Perlmutter, D.C. Circuit, 18 March 2025; certiorari denied 2 March 2026, on the human-authorship requirement for copyright registration (slip opinion and Copyright Office guidance)

• Lehrman v. Lovo, Inc., No. 23-cv-08269 (S.D.N.Y. 10 July 2025)

• Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988)

• Waits v. Frito-Lay, Inc., 978 F.2d 1093 (9th Cir. 1992), for the $375,000 compensatory and $2,000,000 punitive award on voice misappropriation, the vacated Lanham Act damages and the disposition

• Tennessee ELVIS Act, signed 21 March 2024, effective 1 July 2024, and Holland & Knight's analysis of what it added to Tennessee's 1984 Personal Rights Protection Act

• California AB 2602, digital replicas in personal or professional services contracts

• California Civil Code section 3344, which has listed voice among the protected attributes since the 1984 amendment

• NO FAKES Act of 2026, S.4591

• EDPB Opinion 28/2024 on data protection aspects of AI models, 17 December 2024

• GDPR Article 20, right to data portability

• OAIC guidance on biometric scanning under the Privacy Act 1988

• FTC weighs in on the 23andMe bankruptcy, letter dated 31 March 2025

• Regeneron's winning bid for 23andMe, May 2025

• Spotify to acquire Sonantic, June 2022

• Veritone acquires VocaliD, 16 June 2022, with Northeastern University on VocaliD's Human Voicebank and how donor recordings were blended into a recipient's voice

• Descript acquires Lyrebird, 2019

• Coqui shutdown announcement, January 2024

• Voiceprints, AI, and BIPA: New Trends in Biometric Privacy Litigation, American Bar Association

• Loevy + Loevy announcement of the nine Illinois voiceprint class actions, 14 May 2026, for the defendant list and the plaintiffs' professions

• Coqui XTTS v2 model card and the TTS release history, for the November 2023 open release preceding the January 2024 shutdown

• Xu et al., "Machine Unlearning: Solutions and Challenges", for the exact versus approximate unlearning distinction

• Obar, J. A. and Oeldorf-Hirsch, A., "The biggest lie on the Internet", Information, Communication & Society 23(1), 2020, 128 to 147

• European Data Protection Supervisor on machine unlearning