The
MP3 compression code utilizes differences in the ability of the
human ear to hear at different
frequencies to drop
data at certain
frequencies just enough so that the
listener cannot tell the
difference between the
original and the
compressed music. If you put a
dog in however, it will be able to tell the
difference.
The sound signal is compressed into 32 sub-bands through a combination of filtering and down-sampling. Each band corresponds to around 690Hz in bandwidth and has a sample rate of 1.38kHz. A discrete consine transform (very similar to a Fourier transform) is performed on each band.
Each band is then quantized seperately depending on the frequency domain of the band. More data is used in bands in which the psychoacoustic model says the sound is most audible, or more sensitive to the humar ear. Less data is used in bands in which the sound is masked.
The result is that aurally prominent sounds to be stored very precisely while others are cut down. Those will still have substantial amplitude but the amount of data is highly reduced. Yet it will still sound the same as before.