The question teams actually ask is not "is voice cloning legal" but "what do we need in place before we ship this". The answer is that no single statute governs voice cloning. A commercial voice project can touch right of publicity, biometric privacy, data protection, contract law, consumer protection and AI transparency rules simultaneously, and which ones apply depends on whose voice it is, where they live, where you operate and what the synthetic voice will say.
What follows is the 2026 landscape for the jurisdictions most voice AI teams operate in, and then the part that does the actual work: what a voice consent document has to contain to survive contact with any of them. This is general information for planning purposes, not legal advice, and the status of pending legislation changes.
The short version
Cloning a voice with the speaker’s informed, specific, documented consent is lawful in essentially every market a Western company operates in. Cloning without it is where the exposure lives. Almost every legal problem in this field is a consent problem wearing a different hat: consent that was never obtained, consent that was obtained for a different purpose, or consent that was too vague about what the voice would be used for.
So the compliance work splits in two. Get the consent document right, once, properly. Then handle the jurisdiction-specific obligations that attach to the output rather than the input, which in practice means disclosure and labelling.
United States: a state-by-state patchwork
Right of publicity, including the old cases that still bite
Most US states recognise a right of publicity, and several name voice explicitly. California Civil Code § 3344 has covered voice alongside name, signature, photograph and likeness for decades.
Two Ninth Circuit cases matter more than their age suggests, because both involved imitation rather than copying. In Midler v. Ford Motor Co. (1988), Ford hired a backup singer to imitate Bette Midler for an advertisement after Midler declined. In Waits v. Frito-Lay (1992), a sound-alike was used for a Doritos advertisement, and Tom Waits was awarded a multi-million dollar verdict. Neither defendant used the artist’s recordings at all. The relevance to voice cloning is direct: a model trained on somebody else’s data that nevertheless produces a recognisable imitation of a distinctive professional voice is squarely in the territory these cases created.
Tennessee’s ELVIS Act
The Ensuring Likeness Voice and Image Security Act took effect on 1 July 2024 and was the first US law written specifically for AI voice cloning. It amended Tennessee’s Personal Rights Protection Act to add voice as a protected personal property right, and it covers both a person’s actual recorded voice and a simulation that is readily identifiable as them. It also reaches the tooling: making available a service or algorithm whose primary purpose is producing a particular person’s voice without authorisation carries liability of its own. Protection applies to the living and to estates.
California AB 2602 and AB 1836
Signed in September 2024, these are the two provisions most likely to invalidate a contract you already have.
AB 2602 makes a contract term permitting the creation and use of a digital replica of a performer’s voice or likeness unenforceable if it does not contain a reasonably specific description of the intended uses, or if the performer was not represented by a lawyer or a union during negotiation. This is the provision that defeats a broad, boilerplate "all media now known or later devised" grant. Breadth is precisely what makes it fail.
AB 1836 prohibits producing or distributing a digital replica of a deceased personality’s voice or likeness in an expressive work without estate consent, with a statutory minimum award of $10,000.
The NO FAKES Act: still a bill
The NO FAKES Act would create a federal digital replication right in voice and visual likeness, transferable and licensable, surviving death and running up to 70 years post-mortem, with a notice-and-takedown process for platforms and carve-outs for news, documentary, biographical, commentary, criticism and parody uses.
As of September 2026 it is not law. The 2026 version, S. 4591, was advanced unanimously by the Senate Judiciary Committee on 18 June 2026 and awaits action by the full Senate. Plan against the state patchwork, and check the current status before relying on anything here, because this one is moving.
Biometric privacy, and a distinction teams get wrong
Illinois’ Biometric Information Privacy Act names "voiceprint" explicitly, requires written notice and release before collection, and carries a private right of action with statutory damages, which is why it generates the litigation it does. Texas and Washington have comparable statutes without the same private right of action.
Under the GDPR the distinction is sharper than most summaries allow. Article 4(14) defines biometric data as data resulting from specific technical processing allowing or confirming the unique identification of a person. A voice recording is personal data always. It becomes special-category biometric data under Article 9 when it is processed for the purpose of uniquely identifying someone.
Training a TTS model to sound like a consenting speaker is generally not processing for unique identification. Building a speaker-verification system is. Teams get this wrong in both directions, either treating ordinary TTS corpora as if they were fingerprints, or treating voiceprint enrolment as ordinary personal data. Identify which one you are doing before choosing a lawful basis. Our speech data licensing guide covers the contract and records side.
EU AI Act Article 50: now in force
Article 50’s transparency obligations have applied since 2 August 2026, and they attach to the output rather than the training data.
- Providers of systems generating synthetic audio must mark outputs in a machine-readable format detectable as artificially generated. Providers of qualifying synthetic-content systems placed on the market before 2 August 2026 have until 2 December 2026 to take the steps needed to comply with Article 50(2).
- Deployers generating deepfake content, meaning audio resembling a real person, must disclose that it is artificially generated. The Commission’s draft guidelines of 8 May 2026 indicate this applies even without intent to deceive.
For a consented brand voice the practical burden is modest: watermark the output and say somewhere reasonable that the voice is synthetic. It is worth noting that the disclosure duty can apply to audio resembling a real person even where that person consented, so "we had permission" does not by itself discharge it. The EU AI Act and speech data guide covers the obligations attaching to training data and governance.
Elsewhere, briefly
United Kingdom. No standalone right of publicity. Protection runs through passing off, UK GDPR and, for well-known voices, trade mark and advertising rules. The practical driver is UK GDPR plus contract.
Denmark. The Danish government has advanced an amendment to copyright law giving people a copyright-style right over their own likeness and voice against deepfakes, an unusual approach in Europe and one worth watching if you operate in the Nordics.
China. The deep synthesis provisions require consent from the subject whose voice or likeness is synthesised, plus conspicuous labelling of synthetic output. A Beijing Internet Court decision in 2024 found for a voice actor whose recordings were used to build a TTS voice without permission.
Union agreements. Not law, but binding where they apply: SAG-AFTRA’s television, theatrical and interactive media agreements require consent and compensation for digital replicas. If your speakers are union members, the agreement may impose more than the statute does.
What a valid voice consent actually contains
This is the part that does the work. A consent that names these things survives most of the regimes above at once; one that omits them tends to fail several simultaneously.
| Clause | What it has to say | Why |
|---|---|---|
| Synthetic-voice grant | That a model may be trained on the recordings to generate new speech, named explicitly and separately from the recording licence | The single most common gap. A session release does not imply it. |
| Specific intended uses | Products, contexts and categories of content the voice will speak in | California AB 2602 voids terms without a reasonably specific description |
| Prohibited uses | Named exclusions: political content, adult content, endorsements, defamatory speech, whatever the speaker will not agree to | Protects the speaker and is the clause that keeps them willing to come back |
| Territory and term | Where, for how long, and what happens at expiry to both the model and the recordings | Open-ended grants attract scrutiny and are what speakers most often later dispute |
| Compensation | The basis: session fee, usage-based, buyout, or a combination, stated plainly | Union agreements require it and unpaid replicas are the industry’s live grievance |
| Sublicensing and transfer | Whether the voice can move to a parent company, an acquirer, or a customer | Decides what happens to the voice in an acquisition, which nobody thinks about until it happens |
| Withdrawal | What the speaker can revoke, what that obliges you to do, and what is not reversible | GDPR consent is withdrawable; be honest that a trained model cannot always be unbaked |
| Retention and deletion | How long recordings, embeddings and checkpoints are kept, and how deletion is evidenced | Data protection duties attach to the derivatives, not only the WAV files |
| Disclosure | That outputs may be marked and labelled as synthetic | Article 50 requires it; agreeing it up front avoids renegotiation later |
Two failure modes are worth naming because they are so common. Retro-fitting: collecting first and papering it afterwards does not cure the original processing, and a speaker who has since become unreachable cannot sign anything. Consent inherited through a vendor: if a supplier collected the audio, the grant has to reach you, and it has to cover what you specifically intend to do. Ask to see the actual speaker-facing document, not a summary of it. The buying AI training data guide covers that diligence more broadly.
A pre-launch checklist
- Signed consent naming the synthetic-voice grant, specific uses, prohibited uses, territory, term and compensation, for every speaker in the training set.
- A documented lawful basis, and a decision on record about whether Article 9 applies to what you are building.
- Output watermarking and a disclosure surface, if you serve EU users.
- A retention and deletion schedule covering recordings, embeddings and model checkpoints.
- Evidence of the source chain for any licensed or vendor-supplied audio, including the speaker-facing consent text.
- A named owner for re-checking status of pending legislation, NO FAKES foremost.
