|
Abstract : |
A novel approach to speaker identification based on Vladimir Vapnik's support vector network classifiers is presented. Current speaker identification systems are either based on Bayes decision classifiers where speaker densities are parameterized, perhaps by Gaussian mixture models, or on neural net discriminators. Both these approaches use some form of cross-validation to estimate the numbers of parameters and to avoid over training with limited amounts of training data. What sets the support vector network technique apart from all other classifiers is the built in ability of the theory to automatically select a classifier from classes of functions with virtually unlimited numbers of parameters. In addition to reviewing the theory and efficient implementation of support vector networks, issues encountered in applying the technique to speaker identification such as feature selection, channel normalization and computational issues are discussed. Experimental results on a standard telephone conversational speech test are presented. 1, |
