The main reason for using a psychoacoustic model for audio compression
is that given a masking threshold 
, the
amplitude at that frequency may be quantized with a step size proportional
to .
The quantization can be seen as introduction of noise with power
proportional to :
The quantization error can then easily be adjusted to be lower than the masking threshold, and thus become inaudible.
In the implementation of the coder, the psychoacoustic model is adjusted
using only a quantizer (i.e no coding) with step size on
every transform coefficient. The model is adjusted until the effects of
the quantization is inaudable. After this process is finished, the
actual audio coding and quantization is developed.
This way, the coding is kept separate from the psychoacoustic
model. Thus, when I start to design the coder, I can be certain to
get perceptually perfect data independent of coding method.