Part of me wonders if the next step with this is some sort of DL model. I wonder if, trained on one set of melodies (defined in the intuitive sense) it would generate existing copyrighted melodies not in the training set.