It's currently identifying songs by meta-data only. There is an audio fingerprinter (using Last.fm's service) in the works, but as you can imagine that only solves part of the problem: getting fingerprints for data from all the various supported sources is a lot harder.