AISongMakerLab AI music generator playbooks

AI Cover Versions: Where Voice Swapping Crosses the Line

A cover version crosses the line when a listener would believe a specific person performed a song that person never recorded. The line is not about the tool, the model, or how convincing the output is — it is about three separate rights stacked on top of each other: the composition you are covering, the recording you are replacing, and the person whose voice you borrowed. The composition and the recording have long, well-worn routes to permission. The voice does not, because a voice is treated as part of a person's identity rather than as an instrument you can license with a form. Below: a three-layer rights test, a four-band voice scale that tells you which band you are in before you publish, a five-question release gate, and a copy-paste provenance note so the decision is recorded rather than remembered.

Two microphone silhouettes facing each other across a dividing line, one side with smooth blue waves and one side with jagged amber waves

Published 29 October 2026 · about 14 minutes · no signup required to run any test here

The line, in one sentence

Everything below is a restatement of that sentence, but it is worth being precise about why the sentence works. Two things can be true at once: a track can be entirely lawful and still be a cover, and a track can be entirely original and still be a problem.

The practical consequence is that you cannot clear a cover version with one email. You have to clear three layers, and the third layer is the one that does not have a standard form attached to it.

Layer one, layer two, layer three

Most arguments about AI covers collapse because people are arguing about different layers at the same time. Separate them first, then each layer has a straightforward answer.

LayerWhat it protectsWhere permission comes fromWhat a clean cover looks like
1. The compositionThe song itself: words, melody, chord structureThe songwriter or publisher, through a licenceYou record your own performance of a song you are licensed to use
2. The recordingA particular master: the exact sound of a particular takeThe label or recording ownerYou use the original master, or you replace it and hold the licence for your replacement
3. The voiceThe person: their name, likeness, and in many places their voice as a protected attributeThe person themselves, and in some places a specific written or statutory regimeThe performance is attributable to you or to a fictional performer, not to a real individual

Most problems are layer three problems wearing a layer one costume. A song from sixty years ago, in the public domain, with a vocal that sounds exactly like a living singer, is a layer three problem. The age of the composition does not touch the third layer at all, which is exactly why "it is an old song" is such a common and such a bad defence.

Three stacked layers: sheet music paper with staves, a vinyl record with grooves, and a human head-and-shoulders silhouette

Layer one: the composition

To cover a composition you need permission from whoever owns the song. In most markets that permission arrives in a familiar shape: a mechanical licence for the composition, which covers the making of a new recording of a song you did not write, plus whatever arrangement rights apply to how you changed it. You also owe a share of the revenue, which is why cover projects have a business shape as well as a legal one.

Two practical notes that save time later:

Layer two: the recording

Layer two only bites if you use the original master. If you re-record the song yourself — which is what an AI cover version does by definition — there is no original master in your file, so layer two is satisfied by construction. The failure mode here is subtler: shipping the original master's audio inside a new generated track, or using the original master's isolated vocal and calling it a cover.

The second of those is the one that catches people out. Pulling a vocal from an existing record and running it through a voice conversion tool produces output that is a derivative of that recording, whatever the tool did to it. That is a layer two problem and a layer three problem at the same time, and it is the version of this practice that gets accounts claimed rather than politely complained about.

Layer three: the voice

This is the layer with the least predictable process, so it gets the most careful treatment here. Three things are true at once about voice and identity, and you need all three in your head:

  1. Voice and likeness protections exist in many places. A number of jurisdictions have enacted statutes specifically about voice and digital replicas, and the details differ by state or country. Some protect commercial use only, some protect any use, and some require consent that cannot be bought. Read the rules for every territory you distribute in.
  2. Consent to use a recording is not consent to use a voice. A guest vocalist who agreed to sing on your track has not agreed to be a voice model for the next hundred tracks. If you build a model from their audio, you need a separate agreement that says so, in writing, with a stated scope.
  3. Money can buy permission, but it cannot always buy it. For a living voice used commercially, a written agreement with the person is the normal route. For some uses — political advertising, impersonation in certain categories, and a few others that vary by place — permission cannot be sold at all.

None of this requires the tool to be illegal or the output to be fraudulent. It requires you to decide, before release, which of the two sides of the line you are on, and to be able to show how you decided.

The four-band voice scale

The single most useful thing you can build for yourself is a scale with four bands, so you can place a project before you publish it rather than after a complaint arrives. The bands are about what the listener would conclude, not about which technology did what.

BandWhat the listener concludesHow you describe it in the briefRisk posture
A. Genre and era"This is a song from a period and a style"60s soul ballad, warm analog production, close-harmony trioLow. Describe the record, not the performer
B. Archetype"This is a rough, untrained, street-style singer"raw lo-fi bedroom vocalist, off-mic phrasing, unpitched entryLow to moderate. No real individual is implied
C. Sound-alike"That sounds a lot like someone"named-adjacent descriptors only; never a person's nameModerate. Risky if the description is deliberately precise
D. Cloned identity"That is them"a voice model built from a specific person's audio, with their name in the titleHigh. Requires explicit written permission, and in some places cannot be licensed at all

Two failure modes sit at the edges of this scale and are worth naming separately.

The five-question release gate

Run these in order before you publish anything that contains a voice. If a question has an awkward answer, stop at that question — the later ones are cheaper to fix than a takedown.

#QuestionIf yesWhat it triggers
1Is the target voice a specific, living, identifiable real person?Person in questionYou are in band C or D. Stop and get the answer to question 2 in writing
2Do you have a written permission that explicitly covers voice use, model training, and the territories you are publishing in?Scope checkPermission that covers recording only is not enough. Scope it properly or leave band D
3Does the finished release carry that person's name, likeness, or the word "featuring" anywhere?Marketing checkTitle and artwork create their own layer three problem, separate from the audio
4Could a listener reasonably believe this person performed it, and would that belief change anything material — money, prestige, or a false endorsement?Deception checkMaterial deception is the part that attracts enforcement. Band D or a clear labelled performance
5Is the underlying composition cleared for the way you are using it?Layer one checkKeep going only when this is settled; it is the layer people are most often already handling

Notice that questions 1 through 4 concern the voice, not the tool. Swapping tools does not change a single answer in this table. That is worth knowing before you spend money on a better model.

Covering a song versus replacing a singer

Two projects often get confused for one another because both involve "a cover." They have different answers, so it is worth writing down which one you are doing.

What you changedWhat you keptWhat you needWhat the listener thinks
New performance of a licensed composition, original voiceWords and melodyComposition licence, revenue share"A new artist covered this"
New performance of a licensed composition, band D voice with permissionWords, melody, a named person's soundComposition licence plus a voice permission with a stated scope"A collaboration, if it is labelled honestly"
New performance of a licensed composition, band C sound-alikeWords, melody, a resemblanceComposition licence, and a check that the resemblance is not deliberately precise"Sounds like that era or that kind of voice"
Same composition, same master, new vocal treatmentEverything except the vocalPermission for the master, which is a different and harder permission"The original with a new vocal", which is the most common accusation
Everything generated, including the songNothingOriginality of the composition, plus a clean layer three answer"An original track"

The fourth row is the one that ends careers of small channels, and it has a technical signature: the instrumental is recognisably the original master. Generative tools cannot produce that from a description. If your instrumental is a rendered copy of a record you did not license, you have made a derivative recording, whatever the vocal is doing.

A provenance note you can paste

Decisions that live only in your head cannot be defended later, and a note you wrote at the time is far more credible than a reconstruction. Keep one record per released track. Copy the block below and fill it in before the track goes out, not after.

track:____________________________
release date:________________________

composition: [ ] written by me
             [ ] licensed - licensor:__________ date:________
             [ ] public domain in my territories - basis:________
recording:   [ ] performed by me / paid session
             [ ] generated by me, from my own prompt and stems
             [ ] contains any third-party master audio - NO

voice band:  A genre / B archetype / C sound-alike / D cloned identity
             target person:____________________________ (blank for A and B)
permission:  written scope attached? [ ] yes  [ ] no
             covers: [ ] voice use  [ ] model training  [ ] territories:______
             covers: [ ] commercial use  [ ] name/likeness in promotion
release:     real person named in title or artwork? [ ] yes  [ ] no
             labelled as AI-generated in the metadata? [ ] yes  [ ] no
             distribution platform terms checked? [ ] yes  [ ] no
             content-ID or fingerprint registration done? [ ] yes  [ ] no

decision: [ ] publish as-is
          [ ] change band and regenerate
          [ ] seek written permission
          [ ] rename the release and re-check questions 3 and 4
          [ ] do not publish
reason:______________________________________________________
date:____________________

Two fields in that block are the ones people skip and regret skipping: territories and name/likeness in promotion. Permission that covers one country does not cover another, and a name in the title is a separate use from a voice in the file.

Fifteen-minute pre-release check

Before you press publish, work through this once. It is designed to be done on a phone between other things, and it covers the failures that actually happen.

When the right answer is to ask, or to hire

Some projects are worth doing properly and have no clean shortcut. Two cases in particular.

A named artist in a campaign. If money is involved, the artist is a participant rather than a dataset, and the project needs an actual agreement. The cost of that agreement is a known quantity; the cost of a claim is not, and it includes the takedown, the platform record, and the relationship.

A voice you want to keep using. If a voice becomes a recurring character across many tracks, build it with a consenting singer, record it as a session, and own the session agreement outright. That is a different production model from generating a throwaway vocal, and it is the model that survives the second release, which is the release that makes it worth doing.

Common questions

Is it illegal to make an AI cover of a song with a voice that sounds like a famous singer?

Illegal is doing a lot of work in that sentence, and the answer depends on the jurisdiction. What is generally true across jurisdictions is that there are at least two separate questions: whether the composition was cleared, and whether the voice is protected. A cleared composition with an uncleared voice is still an exposure, and it is the exposure people take by accident. If you need a legal answer for a specific release, this page cannot give it to you — the rules differ by place and change. The gate above is a way to organise your own facts before you go looking for that answer.

What if the singer is deceased?

Deceased performers are still protected in many places, and the rules around posthumous use differ from the rules around living performers. Some regimes allow use with permission from an estate, some restrict commercial endorsement, and some have specific carve-outs for satire, parody, documentary and archival use. Do not read "the person is not alive" as "there is no layer three". It is the layer that most often gets skipped, because there is no one to email.

Is a cover version in the public domain still a problem?

Yes, and this is the most common misunderstanding in the whole area. Public domain applies to the composition. It says nothing about a voice, and it says nothing about the recording, which is protected separately and on its own timeline. An old song rendered in the unmistakable voice of a living person is layer three all the way down.

How careful do I have to be with the prompt wording?

Careful enough that the finished release does not identify a person. The wording in your brief is evidence of intent, but it is not the test — the test is what a listener concludes from the finished package. In practice that means two habits: never name the target person in a brief you might have to show, and never rely on vagueness alone, because a precise description plus a matching name in the title is a stronger claim than the name alone.

Does it matter if I do not earn money from it?

Money changes the gravity of a case but is not the switch that turns something into a problem. Several jurisdictions protect voice and likeness outside commercial contexts as well. If your use is genuinely non-commercial and genuinely not misleading, the exposure is lower, but it is not zero, and "I am not making money" is a fact that is hard to establish later if the track took off.

What about the distribution platform's rules?

Read them, and read them for the plan you are on, because the rules usually differ between free and paid tiers and change without much notice. Platform terms are a separate layer from law: a platform can refuse or remove a track for reasons that are perfectly lawful, and their automated checks often key on metadata and artist-name matching rather than on how the audio was made. Filling in the AI disclosure field accurately is cheaper than the alternative.

What if the tool that made the voice is legal to use?

The legality of the tool does not transfer to the output. A tool can be sold and permitted while particular uses of its output are restricted, and your responsibility does not disappear because the vendor sold to everyone. The decision you have to defend is always yours: whose voice, on whose terms, under which territory.

Do I need to register the track with a content ID system?

If you intend to earn from the track, or if you intend to protect it from being used elsewhere, registration is worth doing — and the registration you file is a public claim that you are the rights holder, so it should be accurate. If you covered a composition under a licence, check whether that licence expects you to share registration revenue. Where a track contains material you do not own outright, the registration question and the provenance question are the same question.

What is the safest way to make a cover version?

Write or license the composition properly, use your own voice or a clearly fictional one, name no real person anywhere in the release, and keep the provenance note. That version is boring, cheap, and defensible — which is exactly why it is the version that survives contact with a platform's automated checks.

Can I fix a track after release if I got the band wrong?

Sometimes, and the fix is usually to rename and re-file rather than to regenerate the audio, because the band was determined by the package as much as by the file. A title change can remove the endorsement implication without touching a single sample. What you cannot fix retroactively is the impression already formed by listeners, so the check has to happen before release.

Bottom line: the line is not where the technology stops and the law starts — it is where a listener stops hearing a style and starts hearing a person. Place the project on the four-band scale, run the five-question gate, write the provenance note, and you will know which side of that line you are on before anyone else has to ask you.