The MDCT coefficients are predicted from the two preceding frames, using a separate LMS-adapted (Least Mean Square) predictor for every frequency band. This improves coding efficiency for stationary signals. Residuals after the prediction are non-uniformly quantized and coded using one of 12 different Huffman codes.
In MPEG-2 AAC there are a lot of optional extra features. One of the most interesting is Temporal Noise Shaping (TNS), which works well for transient signals. The idea is, that a tonal signal in the time domain has transient peaks in the frequency domain. The dual of this, is that a signal which is transient in the time domain is ``tonal'' in the frequency domain, i.e consists mainly of a few sines. ``Tonal'' sounds are easily predicted using a LPC approach. Thus, a simple linear predictor is used to predict the next spectral sample (going from low frequencies to high) from it lower-frequency neighbors.