Very interesting. What if you took a large set of musical notes on various instruments, with about a two-second duration, with the same pitch, same volume, and same sample rate, and did the same thing? Eigensounds! And that's just with a time-domain representation of the sound -- you'd get different eigensounds if you used a frequency-domain representation.

I have to wonder if something like this is used by those song-recognizing apps that can tell you what a song is just by hearing a snippet through your phone.