Mediamined is a project I've been working on together with the great folks at Imagine Research for about a year now. With some help from the U.S. National Science Foundation, we're now making public some of our technology.
Mediamined offers the ability to classify sound files into a taxonomy of text labels by analysing the audio signal. We use machine learning techniques to learn the acoustic signatures that correspond to what a bass guitar, male vocals or door slam sounds like, and then classify new sounds according to those signatures. Another capability is to use those signatures to determine acoustic similarity between audio files. This can be used for finding music which is most similar to a query sound file, or to find alternative sound effects, as our CEO Jay explains in a video interview.
Some other write ups have appeared including a nice article on Lockergnome.