New approach for similarity

[mkb218]

Right now, for each root segment in the analysis of the target track, we re-sort the set of all segments by the root. This gets ridiculously slow once the library has about 5 tracks.

One approach to make the search space smaller is this:
Pick a few attributes and thresholds. These can be static or dynamic based on the root (three largest pitch or timbre members, let's say). We filter the search space before sorting, by these criteria

As the number of segments grows, the thresholds get tighter.

[mkb218]

Bucketing doesn't work well. Need something different.

[mkb218]

This sounds like a job for mapreduce. that puts this to 2.0