It is generally fair use if you are:
* not making money from whatever you do
* using software to enhance end results in ways that don't modify the product itself
* using something only for educational purposes
* just analyzing things
However, if you use part of something to make something else, then it's derivative work. If you sell what you make, the copyright gets extremely messy. This is increasingly problematic in a cut-and-paste world where less and less is original. What's happening is that digital natives don't give a flying fuck about copyright and we're likely to wind up back in a phase where people don't even bother to put their names on things because who cares. As an archivist, I hate that, but I do know how to recognize when it's winding up again.
The article also didn't mention the use I'd be most interested in: crunching an entire database of music to derive clusters and patterns. What are the most popular time signatures, rhythms, guitar riffs, or words? You could aim for those in creating new material, which would not be derivative but merely informed. What things are you not hearing, or not hearing much of? Those are areas of maximum potential for innovation. This is basically what I do when I mine fanfic for inspiration. I read so damn fast that I can simple consume vast quantities of material, compiled and correlate, and then load the most desirable elements into my original canons. A quick-and-dirty method would be to read the numbers on the tags in any archive that counts its tags.
It would also be interesting to analyze genres and subgenres of music in search of objective rather than subjective identifiers. Is there really a concrete difference between certain clusters, or is it just something people made up that isn't backed by the content itself? Interesting topic.