Music consists of many different components which are not all audible in the same way. For example, a gentle flute may not be audible if a trumpet is sounded at the same time. Of course, the flute is still present but the listener is simply not able to perceive it at the given moment: in short, the flute is masked by the trumpet.
Many characteristics of human auditory perception are utilized by perceptual audio codecs. Parts of the music which are well-perceived are represented very precisely, while other parts that are not very audible can be represented with lower accuracy. Inaudible information will be discarded. In our example, the trumpet will be represented with great precision and the flute considerably less so. This flexible way of representation helps to reduce the amount of information to be transmitted or stored (file size) and, at the same time, introduces an error (noise) signal. Ideally, this so-called coding noise is masked similarly to the flute signal in the previous example. The smaller the bit-rate of perceptual audio codecs, the less accurately the overall music signal can be represented. Beyond a certain limit (i.e. at very low data rates), the introduced coding noise no longer remains hidden (masked) from the human listener.
Source :
Fraunhofer IIS Homepage