Speaker Recognition: a proposal for a meta-analysis based on hierarchical combination of classifiers

In practical applications of pattern recognition, there are often different features extracted from raw data used for identification. Methods of combining multiple classifiers with different features are viewed as a general problem in various application areas. In this paper some statistical methods based on mixtures of experts models able to combine classifiers and different feature sets are illustrated and applied to the text-independent speaker recognition problem. The comparison takes place on the basis of a dataset composed of more then 50 vocal signals (Italian speakers, phone calls recording).