How to make an AI cover song with any voice?
To make an AI cover with any voice, get or build a voice model for that voice, split the target song into vocals and instrumental, run the vocal through the voice model (RVC voice conversion), then remix. Building a voice needs about 10-30 minutes of clean audio. Using a real person's voice legally requires their consent.
Why — the first-principles explanation
The phrase 'any voice' hides a legal fork you should understand up front. Technically, an AI can convert a vocal into almost any voice, because voice conversion just swaps the timbre while keeping the melody. Practically, 'any voice' splits into three cases: a voice you own (yours), a voice with permission, and a voice you took without consent.
The engineering is the same in all three. You need a voice model — a trained embedding of the target voice. If a preset exists in your tool's library, you can use it directly. If not, you build one by feeding the model 10-30 minutes of clean recordings of that voice. Then you isolate the vocal from the song you want to cover and run it through the model so the tune stays but the voice changes.
The legal weight sits entirely on the voice choice. A person's voice is protected in a growing number of U.S. states by right-of-publicity laws — Tennessee's ELVIS Act (effective July 1, 2024) was the first to explicitly cover AI voice clones. Cloning a celebrity or any identifiable person without consent can expose you to a right-of-publicity claim, even though the voice itself is not copyrighted. So 'any voice' is a technical yes and a legal 'only with permission.'
An example that makes it click
It's like a costume shop. You can dress up as almost anyone, but there are three racks. The first rack is your own clothes — always fine. The second is clothes a friend lends you with a note saying 'go ahead.' The third is a stranger's jacket you grabbed off their chair. The tailoring (making the outfit fit) is identical for all three, but only the first two won't get you in trouble. An AI cover works the same way: the tech makes any voice fit the song, but permission decides which rack you're allowed to shop.
How to do it
- Decide whose voice to use — your own, a licensed preset, or someone who gave clear consent.
- Obtain a voice model: pick a preset from your tool's library, or train one on 10-30 minutes of clean audio.
- Pick the song to cover and separate it into a vocal stem and an instrumental stem.
- Run the isolated vocal through the voice model to convert its timbre while keeping the melody.
- Mix the converted vocal back over the instrumental and adjust levels.
- Confirm you have rights for both the voice (right of publicity) and the song (mechanical/sync license) before publishing.
Key facts
- Voice conversion keeps the original pitch and lyrics and only swaps the voice's timbre.
- Building a custom voice model typically needs 10-30 minutes of clean audio.
- Tennessee's ELVIS Act (effective July 1, 2024) is the first U.S. law explicitly protecting a person's voice from AI cloning.
- A voice itself is not protected by copyright, but using an identifiable voice can trigger right-of-publicity claims.
- Publishing a cover also requires a license for the underlying song's composition.
▶ The 60-second explainer (script)
Here's how to make an AI cover with any voice. Start with a voice model — a trained version of the voice you want. Many tools have a library of ready-made voices, or you can build your own from about ten to thirty minutes of clean audio. Next, take the song you want to cover and split it into vocals and instrumental. Run only the vocal through the voice model. The AI keeps the melody and words exactly, but changes the voice. Then mix the new vocal back over the instrumental and you're done. One important warning: technically the AI can copy almost any voice, but legally a real person's voice is protected in states like Tennessee under the ELVIS Act. Use your own voice, a licensed preset, or one you have clear permission to use.
What authoritative sources say
People also ask
Can the AI really use any voice?
Technically almost any, because it only swaps timbre. Legally you should stick to your own voice, licensed presets, or voices you have permission to use.
Do I need the person's permission?
Yes for any real, identifiable person. Right-of-publicity laws like Tennessee's ELVIS Act protect a person's voice from unauthorized AI cloning.
How long does building a voice model take?
The audio requirement is usually 10-30 minutes of clean recordings; training itself often takes minutes on hosted tools.
Is the melody generated by AI?
No. For covers, the AI keeps the original melody and lyrics and only changes the voice.