Well, I think that the presentation of the file on file exchange says a lot about how it works. It looks like they transform the sound to frequency domain, then remove the characteristic vuvuzela frequencies, and finally transform back.
However, the title of the file submission is somewhat missleading. The vuvuzela sound isn't "noise" per se, but a series of fairly distinct frequencies that can be subtracted. That means that the algorithm cannot remove traditional noise that is chaotic in nature.
