The first thing is to define what is a «good» and a «bad» sound. This is an extremely tricky issue, since the networks need numeric inputs. And music is whole bunch of numbers.
I know from people doing research in identifying how similar two sounds are, and imitation, say: you hear a sound and try to make another that sounds like it. Like when you hum a song or similar. That is by no means easy. These guys are using something similar to feature extraction, with Fourier transforms and energy and such things. They feed the networks with the (selected) features and... Train.
Now, to return to your original question: *What do you present as target during training?* You can present different types of music as categories and classify (I couldn't help but think on this research with fish). Or you define categories of music you like and see if the network can classify them ;)
One basic decision here is how long you get a piece of sound. Since it is needed to analyse frequency, this is a key issue. Since you talked about DNN, I was wondering if you wanted to do it online, as a stream, in which case I don't have the slightest idea where to begin, other than do it after a little while.
Other idea: I remember a little sketch in this series about a researcher that makes use of the relations between peaks in the Fourier spectrum in order to differentiate noise from music.