Besides the complications already mentioned, it is important to remember 
that human sound source separation is greatly enhanced by the shape of 
the ears combined with slight movements of the head.

If you sit at a party situation, you can easily focus on different 
conversations. But if you do a high quality stereo recording from your 
position and then listen to that, you will find it far more difficult 
to do the same thing. So a FFT/AI/whatever working from an MP3 starts 
out with a huge disadvantage since so much information has already been 
lost.

-- Jecel