When I googled it, this paper come up near the top:
http://www.ie.itcr.ac.cr/palvarado/papers/TSD2011_Paper_04.pdf
It's mostly theoretical, but gives formulas and block diagrams that are helpful
in understanding the process of AGC. Note that in their application there is
a delay introduced in the signal path in order to allow the energy estimate of
a particular sample to be applied to the same sample. For "live" audio,
where a delay may be undesirable, you would normally work with the energy
estimate of preceding samples. This means that for an application like
the gunshot noise, a typical broadcast style AGC wouldn't work well. However
for speech and music, where you don't typically have such isolated peak
amplitude spikes, you can get away without the delay.
The main idea of AGC is that you first take the envelope of the signal,
then filter it to get an average energy measurement. Then you use
an inverse function of the energy as a gain on the unfiltered signal.
Normally there are limits imposed on the gain function to avoid effects
like line-frequency hum or hissing during quiet periods.
-- Gabor
-- Gabor
