Consider the case where a speech, , is passed through the linear
predictor, *a*, corresponding to a template. The residual mean squared
error, *E*, is:

So can define

Or in the autocorrelation case:

Let be the augmented vector of ``template'' or
``reference'' LP coefficients and the
augmented vector of ``observed'' or
``unknown'' LP coefficients then:

In the autocorrelation case this can be computed with similar complexity
to the Euclidean distance.

Speech Vision Robotics group/Tony Robinson