From [6] I take the approximation of the curve in figure 1 as
where ATH(f) denotes the ATH in dB, and f is the frequency in Hz. The
problem in digital audio coding,
is that it is not known at what absolute level the sound will be played.
One common solution is to set the lowest point on the curve in
(6) to be equal to the sound pressure level of a sine with
amplitude 
LSB.
In the coder, 16-bit samples with normalized amplitude to
are used.
Thus, the smallest possible sine has amplitude 
, and
has a power of 
.
In experiments, though, that setting did not seem to ``keep'' high enough
frequencies for transient sounds, so the model was chosen as
ATH'(f) = ATH(f)-114.
