"A quest by researchers at the University of Rochester resulted in a 20-second clarinet solo being compressed into less than a single kilobyte of data -- nearly 1,000 times smaller than a standard MP3 representing the same audio. In order to get this done, they created a model of the clarinet itself -- essentially replicating each aspect of the sound rather than creating thousands of digital samples from a performance of it. The resulting file occupies less than a kilobyte despite including all of the audio materials. By comparison, the same clarinet sample would occupy 32KB as a MIDI file."
Pretty interesting research, but there are so many variable that I doubt any model would capture all of them. As a result, I don't think this will catch on to replace MP3's for specific musical works, but would be ideal for reproducing other audio data.