Tips on speaker recognition by autoregressive parameters and connectionist methods

摘要

This study reveals more interesting aspects on speaker and speech recognition as: 1. different importance of certain spectral frequency bands on the process of speaker and speech recognition; 2. signal phase has a significant importance; and 3. vowel recognition is preponderant in the decision weighting. To resolve the paradox described in A.J. Grichnik (2000), autoregressive (AR) coefficients were used to compute feature vectors in order to teach neural networks (NN). Tests made by using a two layer perceptron (MLP) were compared to a radial basis function (RBF) network in order to obtain the best recognition results.