While I was trying to figure out a way to approximate an album gain from track gains and track durations alone, Garf came up with the idea of using the 95th percentile of a sorted list of all track gains, weighted by track durations (track gain values in that list are repeated as many times as the number of seconds of their respective track).
As it turns out, the 34th percentile is the magic number that matches album gains the closest in my personal collection (391 albums): 0.25 dB deviation on average, with a maximum of 1.3 dB, including odd cases with silent tracks and wild variations between tracks.
Any idea why that number works so well?
Most collections should have an individual number like that.
Well, I don't know about that. What's interesting here is that it matches even albums in my collection that are very different from the rest: parts of silence, tracks gains varying greatly, etc…