Found a dumb mistake in my feature extractor a couple weeks ago. I finally bit the bullet last week and fixed it, rescaned all my data (100GB!), and retrained my classifiers. Got a 10% boost in accuracy, so it's not only more correct, but useful too, which is a nice bonus. My simple linear network can distinguish between techno and pop music with 91% accuracy, and it's only using spectral features. This is going to rock once I add tempo, beat strength, and rhythm signatures. I'm in a feature freeze until I submit the thesis though. Must... document...!