The idea in one sentence

An AI cover is a new recording of an existing song in which software, not a human performer, supplies the voice, the arrangement or both.

Two very different technologies

Voice conversion takes an existing vocal track and re-sings it in another voice. The melody, timing and phrasing come from the original performance; only the timbre changes. This is the technology behind the viral clips of one star "singing" another star's hit.

Generative music models work the other way round. You give a text prompt, lyrics and sometimes a reference clip, and the model produces a whole new recording, instruments included. Several of these tools have a cover or remix feature that keeps a song's melody and changes its genre, language or words.

What makes a cover sound good

  • A clean source. Isolated vocals convert far better than a full mix.
  • A voice that suits the range. A deep voice pushed two octaves up sounds robotic.
  • Lyrics written to be sung. Count syllables and keep the stresses where the melody puts them.
  • Restraint. One clear change (new genre, new language or new words) beats changing everything at once.

Copyright: the song and the recording are separate

This is general information, not legal advice. A song carries two copyrights: one in the composition (melody and lyrics) and one in the sound recording. A traditional cover re-records the composition, and in many countries a standard licence covers that as long as you don't change the song's basic character.

Changing the lyrics is different. New words on someone else's melody is usually a derivative work, which needs the publisher's permission rather than a standard cover licence. Parody has some protection in some countries, but it is a narrow defence and not a general pass.

Using the original recording itself, for example by converting the real vocal track, involves the recording copyright too.

Voices are protected as well

Copyright doesn't cover a voice, but other laws do. Many places recognise a person's right to control commercial use of their likeness, and that increasingly includes their voice. Tennessee's ELVIS Act, passed in 2024, names voice explicitly, and other jurisdictions have followed with their own rules on digital replicas.

The practical lesson came in 2023, when a track imitating two well-known artists spread across streaming services and was taken down within days. Platforms now remove cloned-voice tracks quickly when rights holders complain.

Where the industry stands

Major record labels sued two of the largest AI music generators in 2024 over the recordings used to train them. Since then the direction has shifted from lawsuits toward licensing, with labels and AI music companies announcing deals for licensed models. Expect rules on downloads and commercial use to keep changing, and check a tool's current terms before you publish anything made with it.

How to make an AI cover without trouble

  • Use your own voice, a voice you have permission to use, or the tool's stock voices.
  • Prefer songs in the public domain, or get a licence for the composition.
  • Label it as AI-generated when you post it. Most platforms now ask for that.
  • Keep fan projects non-commercial, and take them down if a rights holder asks.
  • Never present a cloned voice as the real artist.

Writing the new lyrics

The hard part of a rewritten cover is the words, and that is where a chat model helps. Paste the original's syllable pattern, describe the new theme, and ask for lyrics that keep the same stresses and rhyme scheme. An uncensored model is useful when the song is dark, crude or satirical, since mainstream assistants often refuse lyrics with swearing or adult themes.

In UncutGPT you can try the same request across several models and keep the best lines from each.