Clustering of Songs (The Joe Walsh Problem)

Question

The Eagles are a rock supergroup from the 70s and 80s, responsible for such classics as Hotel California. They have two quite distinctive sounds, one where guitarist Joe Walsh is present (for example, in Life in the Fast Lane) and one where he is absent. The latter songs have a markedly more sombre/boring feel.

I'm curious to understand the degree to which an (unsupervised) learning algorithm would be able to detect the difference between the two sounds. One could imagine that it would be easy to tell the difference between speed metal and classical music, but what about sounds by the same band.

How would I set up such an experiment? Assume that I already have the relevant audio files in some standard format.

Note that this should also apply to other rock groups, such as AC/DC who had a change of lead singer in 1980, and possibly even to other genres, possibly even more modern music.

Answer 1

What you want to do seems to be known as Audio Feature Extraction or, more specifically, Music Information Retrieval, that is automated methods that distill characteristics out of (sets of) music files. You would have to extract features of samples of both equivalence classes and look at differences that can inform song choice.

Researchy tools are available for instance here and here. A finalist of Google Science Fair 2011 presented interesting stuff but I can not find his tools anywhere.