We quickly see the DPCM, because LPC really is a combination of the DPCM kind of coding, as well as the vocoder kind of coding. DPCM is based on quantizing the error between the two consecutive samples because the samples are correlated. And their difference, therefore, will have a smaller variance and will require a smaller number of quantization levels. Instead of the two consecutive samples, what we now do in DPCM is to pick the current speed sample and the predicted speed sample, take their difference, quantize them and send the quantized value onto the line. One can use the instantaneous and the syllabic compending in DPCM again for the sake of better dynamic range.

---