Using an AI model of your own voice on your own song: copyright, disclosure and free tools

Yes, you can sing a guide vocal, run it through an AI model of your own voice and release it on your own song, and the main tool for doing it is free. The catch is paperwork, not permission: in the US you disclose the AI part when you register the copyright, and you declare it through your distributor where the platform asks. Everything here is as we read the sources on 29 Oct 2026; this is not legal advice.

Start free: the tool

RVC (Retrieval-based Voice Conversion) is MIT-licensed and free to download. Its project page says you can "easily train a good VC model with voice data <= 10 mins" and recommends at least 10 minutes of low-noise speech. Its base model was trained on nearly 50 hours of the open VCTK dataset, which the authors give as the reason it can be used without copyright worries. The cost is your time and a capable GPU, not a subscription. Our voice conversion explainer covers how it works.

For a model of your own voice, record those ten minutes yourself, dry and clean. Then you own the training audio, and nobody else's rights sit in it.

Copyright: the Randy Travis precedent

The US Copyright Office's Part 2 report describes registering a Randy Travis recording that was built from James Dupré's sung performance, translated by "a special-purpose AI vocal model ... as a tool" into Travis's voice while keeping Dupré's cadence, phrasing and dynamics. That is exactly the own-voice workflow: a human performance goes in, a different timbre comes out.

What we think: this is the strongest signal yet that voice conversion of a real human take is treated as a tool, not as the author. The song you wrote and the phrasing you sang are still yours. What you should not claim is a vocal generated from text or a prompt with no performance underneath.

Registering it in the US

The Office's registration guidance says applicants have a duty to disclose AI-generated material in a work, must use the Standard Application, should exclude AI-generated content that is more than de minimis, and should not list an AI tool as an author or co-author. If you registered before and left the AI out, the fix is a supplementary registration. Step-by-step form help is in how to register an AI-assisted song and the wider US copyright registration guide.

A sensible description: you wrote the composition and performed the vocal; the timbre was converted with an AI voice model trained on your own recordings. Whether the converted timbre counts as more than de minimis is a judgement call; when in doubt, disclose.

Distributor and platform disclosure

• Spotify. AI credits are added through your distributor when you upload, are in beta with certain distributors only, and show in song credits and Now Playing on mobile, not as track labels. Spotify's AI policy allows vocal impersonation only when the impersonated artist has authorised it. Using your own voice means you are the one authorising.
• YouTube. Music partners declare Fully, Partly or No Gen AI. Leave it blank and YouTube treats the status as unknown and may classify it from its own signals. An own-voice conversion over your own song and production is, on any honest reading, Partly.

Our distributor disclosure guide and the streaming platforms AI policy tracker list where each service stands.

What we think: declare it. Partial disclosure costs you almost nothing today, and letting a platform's detector decide for you is the worse outcome.

Keep the model to yourself

A model of your voice is a copy of your voice. Store it privately and do not upload it to public model-sharing sites unless you mean to license it; see protect your voice from AI cloning and license your voice for AI.