Deep Learning Machine Solves the Cocktail Party Problem

Audio samples of the results would be nice, but I didn't find any links in the pdf

My guess would be that extracting useful audio is possible, but the idea evoked by the paper's title "Deep Karaoke", of flawlessly removing vocals from a finished song, is a lot harder, or maybe impossible. Anyway this has many useful applications