So even after equalizing, all the top components are similar variants of kick drums. I think this is a problem - PCA should really identify maximally different

So even after equalizing, all the top components are similar variants of kick drums. I think this is a problem - PCA should really identify maximally different things for the top components. If they're not perceptually different, then the PCA is not doing the best job at capturing perceptually relevant variance. I presume what is happening is that the drums all have slight variants in timings, and PCA is not phase insensitive, so sees them as totally different. I wonder if you'd get better results if you downsampled even more agressively, or did some more sophisticated feature transformation.