Recently, Google researchers announced a new generative Artificial Intelligence service: MusicLM. The subject in question is able to generate music starting from "rich descriptive captions", and can transform a mumbled melody into an effective piece, even of different genres.
The service has been "trained" though the use of a database of unlabelled music, together with a sustainable amount of sound descriptions extrapolated from MusicCaps, a new dataset made up of thousands of "song - description" pairs.
The idea of ​​generating music via AI isn't exactly new, but previously the composed melody was then further processed by human hand; with MusicLM this step is not necessary.
The creators of the service have recognized some potential problems that may arise as a result of using MusicLM, and have already stated that they are working to remove these doubts.