Linear prediction analysis

To actually calculate the LPC coefficients, the idea is very simple: for each frame (of 5–10 ms duration) we need to find parameters, a ₁ to a_p such that the error term e[t ] is as small as possible. However, to put this into practise requires a knowledge of mathematics beyond the level of this course (see Schroeder 1985: 55–6 for the details). And in general, we do not need to understand the method in order to use it in practise. My program lpcana.c , for instance, uses the function memcof, taken from Press et al. (1992: 568–9). This takes a vector of data (an array of e.g. 80 consecutive samples), the size of each frame n, and the number of desired coefficients m (i.e. the prediction order), and returns an array d containing the m coefficients. (It also returns the mean square error, though we do not use this, as we have to calculate the sample-by-sample error anyway, in order to obtain the residual.)

Next: Trying it out