I've been toying with ideas like this for a long time now. I think the fitness function is critical, but the problem is that taste is subjective, and you need to listen to many riffs/melodies to evolve to a "good" state. Also, you either start with specific melodies, in which case you would skew the results, or start with random noise, in which case it would take a very long time to evolve to anything good. So it seems like you must constrain i tsomehow, such as "only use 12 tones, with full/half/quarter/third notes".

But anyways, my idea for a way to resolve the problem of fitness taking forever would be to livestream it on Twitch, in the same vein as the "Twitch plays Pokemon" where viewers can input commands to vote for an action, they could vote on the fitness of musical tracks.