What the ELVIS Act protects in a voice
Tennessee made voice a protected personal right and reached the tool layer, not only publishers. Why building with voice clones carries statutory exposure.
Most right-of-publicity statutes stop at the person who published the thing. Tennessee's does not, and that is the detail worth understanding before you build anything on top of a voice.
The Ensuring Likeness, Voice, and Image Security Act — the ELVIS Act, codified at Tenn. Code §47-25-1101 et seq. — took effect on 1 July 2024. Two structural choices in it matter more than the headline. It added voice as a protected personal right alongside name, photograph and likeness. And it made unauthorised digital replicas actionable against those who provide the tooling, not only against whoever posted the clip.
This is general information, not legal advice.
Voice as a right in its own right
Before this, voice sat in an awkward gap. A recording of someone singing is protected by copyright in the recording. A composition is protected separately. The voice itself — the timbre and delivery that make a listener say "that's them" — was protected unevenly, if at all, depending on the state.
The ELVIS Act closed that by amending Tennessee's existing Personal Rights Protection Act to name voice as a protected attribute. That makes voice a first-class subject of the statute rather than something you argue into "likeness" by analogy.
The practical consequence is that the usual copyright reflex gives you the wrong answer. Content ID matches recordings and compositions. It does not match a voice. A synthetic performance that reproduces no protected recording can sail through a copyright clearance process and still be an unauthorised voice replica. YouTube's synthetic singing voice detection extends toward that gap, but it is a platform tool, not a legal safe harbour, and it should not be assumed comprehensive. Our write-up on the hole in likeness tooling covers why voice-first creators are the least protected group in practice.
The "we never said it was them" workaround does not work
The other thing the statute does is reach imitations that merely sound like the person. You do not have to use their name, and you do not have to have used a recording of them, for it to count.
That is the specific defence this was written to remove. "We used a session singer who sounds like her" and "we never claimed it was him" are not exits. If the audience hears the person, the statute is interested. Whether a given performance crosses that line is a fact question, which is exactly why it is expensive to be wrong about.
Why the tooling layer carries exposure
Here is the part that separates the ELVIS Act from ordinary publicity law. Most statutes attach to the use — the ad that ran, the track that shipped. Liability flows to the advertiser and the publisher. The vendor who supplied the software sits outside the chain.
Tennessee's statute makes the tool side actionable too. Map the actors in a typical unauthorised voice clone and the difference is obvious:
| Actor | Exposure under an ordinary publicity statute | Under the ELVIS Act |
|---|---|---|
| The brand whose ad it was | Yes | Yes |
| The producer who generated the clip | Usually | Yes |
| The platform hosting it | Limited, and heavily contested | Contested, plus its own policy duties |
| The service that produced the replica | Generally outside the frame | In the frame |
This changes the risk calculus for anyone in the middle of that stack — not just model vendors, but agencies running an internal cloning service for clients, tool builders wrapping a TTS API, and marketplaces listing voices they did not verify.
The line that will get argued over is the difference between a general-purpose speech model and a service whose point is producing one named person's voice. Those are not the same product and they should not be treated as the same risk. But the argument now happens under a statute that contemplates provider liability, rather than one where the provider was never a defendant to begin with.
What this means if you build with voice
Five things that hold up regardless of how the case law develops.
- Verify consent at enrolment, not at export. The moment that carries exposure is when a voice model is created from someone's recordings. A consent step that happens after generation is documentation, not permission.
- Treat a voice grant as separate from a likeness grant. A face release does not license the voice. If your paperwork has one signature covering "likeness," it probably has a gap.
- Do not ship "sounds like" presets. A named-artist soundalike preset is the clearest version of the thing the statute targets, and it is trivially discoverable.
- Know whose voice is in your library. If you resell or expose a voice marketplace, provenance for each voice is your problem, not the uploader's alone. Our note on why a signed deal is not a cleared model applies the same logic to music.
- Keep the enrolment record with the voice. Who consented, when, to what scope, and for how long. Usage rights covers the vocabulary.
For most production work the exposure is avoidable by construction. Cloning your own voice involves no third-party consent at all, and a stock voiceover library covers a wide range of narration without touching anyone's personal right. The voice cloning entry sets out the baseline rule plainly, and voice fine-tunes vs instant clones covers which method you actually need — a distinction that also changes how much of a person's voice you are retaining.
What the statute does not cover
- It is Tennessee. It binds where it binds. There is still no federal right of publicity in the United States, and the NO FAKES Act, which would create one, advanced out of the Senate Judiciary Committee on 18 June 2026 and remains a bill.
- It is not a disclosure law. Consent does not exempt you from labelling. EU AI Act Article 50 deployer duties apply from 2 August 2026, and New York's synthetic performer disclosure for ads reaching New York consumers took effect 9 June 2026.
- It is not a platform policy. Spotify bans unauthorised voice clones and impersonation on its own terms, independent of any statute. Platform rules bite faster than lawsuits and change without notice.
- It does not settle training. Whether a model provider could lawfully train on a corpus is a separate question being litigated elsewhere. What the ELVIS Act governs is the replica you produced and the tool that produced it.
The broader map of how these regimes interact is in our voice and likeness law overview.
FAQ
Does the ELVIS Act only protect musicians?
No. It amended a general personal-rights statute that already covered name, photograph and likeness, and added voice to that list. Nothing limits it to recording artists — the Nashville framing is about why Tennessee legislated first, not about who is covered.
Does a voice clone have to use the person's name to be covered?
No. The statute reaches an imitation that sounds like the individual, so omitting the name is not a workaround. This is the loophole it was specifically drafted to close.
If I use a general-purpose text-to-speech model, am I a "tool provider"?
Using a tool is not providing one. The provider question attaches to whoever makes the replication capability available. Where it gets genuinely uncertain is a service built or marketed to produce a particular person's voice, which is a different product from general text to speech even when the underlying model is the same.
Does Content ID catch unauthorised voice clones?
No. Content ID matches recordings and compositions. A soundalike that reproduces neither generally passes through it. Voice imitation is a personal-rights question, not a copyright-matching one, which is why clearing a track for copyright tells you nothing about whether the voice was cleared.