This article is from the Audio Professional FAQ, by with numerous contributions by Gabe M. Wiener others.
Yes, CDs are limited to 16 bits, but we can use >16-bit systems to produce
16-bit CDs with higher quality than we could otherwise.
We are able to record audio with effective 20-bit resolution nowadays.
The finest A/D converter systems have THD+N values around -118 dB with
linearity extending far below even that. When it comes time to reduce
our word-length to 16 bits, we can use any one of a variety of noise
shaping curves, the job of which is to mix with our 24-bit audio, shift
the dither spectrum of the noise into areas where our ears are less
sensitive, thus enabling the noise component to comprise audio information
at the spectral areas where our ears are most sensitive. See Lipschitz's
seminal papers for fuller detail on this subject.
Furthermore, we often perform DSP calculations on our audio, and to that
end it is worthwhile to carry out the arithmetic with as much precision
as we can in order to avoid rounding errors. Most digital mixers carry
their math out to 24-bit precision at the I/O, with significantly longer
word lengths internally. As a result, two 16-bit signals mixed together
can produce a valid 24-bit output word. For that matter, a 16-bit signal
subjected to a level change can produce a 24-bit output if desired (except,
of course, for a level change that is a multiple of 6 dB, as that's just
a shift left or right).
The number of noise shaping curves available today is staggering. Sony
SBM, Weiss, Meridian 618, Sonic TBM, Apogee UV-22, Prism SNS, Lexicon
PONS, Waves, and, of course, the classic Lipschitz curve are just a few
of the multitudinous options that now exist. [Gabe]
 
Continue to: