Greg Milner in the New York Times:
I’m talking about loudness as a measure of sound within a particular recording. Our ears perceive loudness in an environment by reflexively noting the dynamic range — the difference between the softest and loudest sounds (in this case, the environment is the recording itself, not the room you are playing it in). A blaring television commercial may make us turn down the volume of our sets, but its sonic peaks are no higher than the regular programming preceding it. The commercial just hits those peaks more often. A radio station playing classical music may be broadcasting a signal with the same maximum strength as one playing hip-hop, but the classical station broadcast will hit that peak perhaps once every few minutes, while the hip-hop station’s signal may peak several times per second.
A loud environment in this sense is one with a limited dynamic range — highs that peak very high, and lows that aren’t much lower. For decades, musicians and engineers have employed dynamic range compression to make recordings sound fuller. Compression boosts the quieter parts and tamps down louder ones to create a narrower range. Historically, compression was usually applied during the mastering stage, the final steps through which a finished recording becomes a commercial release.
More here.