The most successful statistical learning principle used to achieve separation of sound sources is the Independent Component Analysis (ICA). The original setting is based on an assumption that all sources are statistically independent and at most ones source is Gaussian. The independence is obviously doubtful in practice where musical sound sources are often played in unison and people often talk coherently. However, many algorithms give surprisingly good separation (more accurately, good suppression of the interfering sources). Also, even if the sources ar enot truly independent, a meaningful interpretation of ICA as a projection pursuit and sparse coding exist as discussed in [32]. For a more complete list of people working in this area and their papers, Paris Smaragdis' ICA webpage Te-Won Lee's ICA webpage. For a good introduction to ICA and blind source separation (BSS), see Aapo Hyvarinen's ICA