At bit rates of around 16 kbits/s and lower the quality of waveform codecs falls rapidly. Thus, at these rates, CELP codecs and their derivatives, tend to be used. CELP stands for Code Excited Linear Prediction. A code representing a voice pattern is transmitted, rather than coding short samples of voice signal. The code is calculated by looking at the long-term changes in the voice and then predicting how the voice pattern will change. Since the pattern takes some time to establish, CELP codecs produce a delay.
A backward adaptive CELP codec was standardized in 1992 as G728. This codec uses backward adaptation in its calculations. This means that, rather than buffer 20 ms or so of the input speech for its calculations, the pattern of previous speech is used. This means that the G728 codec has to analyse fewer samples than traditional CELP codecs. The G728 codec uses only 5 samples giving it a total delay of less than 2 ms.