I was thinking. Machine learning is probably very important for musicians using digital audio workstation software. With lots of parameters. It should be possible. With the proper error functions. To fine tune parameters of everything in the audio project. For example vsti instrument settings and effects and more.
The error function is the most important part. How can a network tell if the audio sounds good. Here some creativity is needed. Are there a number of smart error functions we can develop? Or maybe its not necessary maybe a network can learn to recognize each type of music and improve as much as possible with the settings.
Another idea with machine learning is to cooperate with the algorithm. If a simple algorithm can predict a tune then a listener can do the same. The idea is that this will make the song more pleasant to listen to. That is. You change the melody to be a little more predictable in the style of the music.