Actually we need to define melodies first. If we expend from Western music theories, the possibility becomes endless (even from information theory perspective).

Actually we need to define melodies first. If we expend from Western music theories, the possibility becomes endless (even from information theory perspective). Also, how long should the melodies be? What would be our perception chunk for melodies? And what if we have multiple layers? See how you can tweak the 'melodies' of Kraftwerk's The Model: https://glicol.org/demo#themodel https://glicol.org/demo#themodel In music AI, there is a famous distinction between symbolic and subsymbolic. I guess when the entropy of symbolic reaches some threshold, switching to subsymbolic raw audio sample generation becomes natural.