Why training data matters
A music model learns everything it knows from its training data. The breadth of that data decides which genres and instruments it handles well; its source decides the legal and ethical questions around the model.
Possible sources
- Licensed catalogues from labels, publishers or production-music libraries, used with permission.
- Commissioned recordings made specifically for training.
- Public domain and openly licensed music, such as CC0 recordings.
- Scraped audio collected from the web without permission, the most contested source.
The legal questions
In June 2024, major record labels, coordinated by the RIAA, sued two prominent AI music companies, alleging their models were trained on copyrighted recordings without permission. Since then some labels have announced settlements and licensing deals with AI music companies. The core legal question, whether training on copyrighted music without permission is lawful, is being argued in courts and policy reviews in several countries and has no single answer yet.
Certification and transparency
Independent schemes such as Fairly Trained certify companies that train only on licensed or permitted data. Some providers publish summaries of their training sources; many do not.
What to ask a provider
- What kinds of data was the model trained on?
- Is any of it licensed, and from whom?
- Does the licence you receive cover your intended use?
- What happens if a generated track closely resembles an existing recording?
Where Sonvela stands
We will publish details about the models behind generation on our about page as they are finalised. We do not claim certifications we do not hold. The tracks in our free library are CC0 recordings by named artists, credited on every page, and are not AI-generated.
Written by the Sonvela editorial team with AI drafting assistance, then checked against the sources linked in the text. Spotted an error? Tell us and we will correct it.