These sound examples were given to me by Tomohiro Nakatani and illustrate the state, in 1996, of the voice-separation system developed in the group of Hiroshi Okuno.
Their "residue-driven" approach uses "agents" to track sets of harmonically-related partials, which are subtracted-away from the time-domain signal to leave a residue, which can then trigger the instantiation of additional agents. Recent developments of their system have incorporated spatial (binaural) information and several approaches for handling aperiodic sound.
There are many more examples available at Nakatani's site.