On the use of features from prediction residual signals in speaker identification