This article is from the Audio Professional FAQ, by with numerous contributions by Gabe M. Wiener others.
Broadband noise is usually better tackled in the frequency domain. What
this entails is taking a block of data (as in the impulsive noise case)
but then calculating its spectrum. From the spectrum an estimate can
be made of which frequencies contain mostly signal, and which contain
mostly noise.
To help in making this discrimination we first take a "fingerprint" of
the noise from an unrecorded section, such as the lead-in groove of a
record, or a silence between movements of a symphony. This spectrum of
this fingerprint is then compared with the spectrum of each block of
musical data in order to decide what is noise and what is music.
The denoising process itself can be thought of as an automagically-
controlled, cut-only graphic equaliser. For each block, the algorithm
adjusts the attenuation of each frequency band so as to let the music
through, but not the noise. If the SNR in a particular band is high
(ie lots of signal, little noise) then the gain is left close to
unity. If the SNR is poor in a given band, then that band is heavily
attenuated. [Chris]
 
Continue to: