At long last I have something new to look at. After a temporarily debilitating deletion of everything on my computer I have been able to pick up the pieces and get some data to look at. What you are seeing above is my first look at the 2nd iteration of the Visual Music Collaborative. It is a map of the sample data from songs in which I used the Databionic Music Miner software to extract audio features from. Features are gathered into twenty values which representing various qualities of the music. Instead of using the built in mapping feature provided by Music Miner, I used the Kohonen algorithm to map the feature vectors into a two-dimensional map. Songs with similar features, ideally, should end up close to each other and sounds that sound different, stray. Each color above represents a different artist and I chose a selection of artists to test the accuracy of the system. The artists represented above are The Beach Boys, Atlas Sound, Beach House, Todd Rundgen, the Vaselines, Blond Redhead, Kraftwerk and Alexander Robotnik. While it’s not totally on point, it is working well enough to group artists together. I was surprised to find that Todd Rundgren and the Atlas Sound are continually paired. Kraftwerk is consistently an outlier (represented by the scattered orange cells above).
After reading the various publications about self organizing maps and audio feature extraction, one of the disclaimers is that accuracy of the maps depend heavily on the type of music used in the system. There is also the issue of perception. Just because it says that Todd Rundgren and Atlas Sound sound similar, doesn’t mean that’s how I think of them. For these reasons, I’ve added twenty sliders that allow you to adjust how heavily each feature is relied upon. In tuning the map, you can watch the artists reposition into a configuration you agree to. It’s exciting to watch.
So, all that’s left to do now is extract audio features from another 20,000 songs and map them! No biggie right. Then I need to re-instill the “community” aspect, letting everyone in the collaborative view the map of the entire community.