A novel approach to speaker verification

The authors report on a novel method for telephone-based speaker verification. In this method, K-means clustering is used during training for robust speaker reference templates. The classification is made through the use of the Fourier-Bessel functions, which transform the original problem to a multidimensional detection problem. This technique allows the combination of multiple feature sets in a single classification test. Experiments with a number of speakers and words over the telephone network show the potential benefits of the new techniques using the standard and multivariate Gaussian classifiers