Two major steps are involved in the general feature
extraction for speaker recognition. First a set of
predictors coefficients are determined by the LPC
analysis and in the second step these coefficients are
transformed into feature vectors.
The basic idea behind linear predictive coding
(LPC) is that a sample of speech can be approxi-
mated as a linear combination of the past ‘p’ speech
samples. By minimizing the square difference be-
tween the actual speech samples and the linearly
predicted ones, one can determine the predictor co-
efficients; i.e., the weighting coefficients of the linear
combination.
41