Score fusion methods for text-independent speaker verification applications

Speaker verification methods are various and use different types of features, but each system alone do not perform satisfactory results. This paper makes a comparison of different features and methods for score fusion for an independent speaker verification application. Several types of spectral features are used as speaker data. The scores obtained with these types of features were fusioned with combination methods (as: mean, sum, max, min, weighted sum) and classification methods (as: SVM, linear discriminant). The experiments were performed on a laboratory registered database for Romanian language and demonstrate that fusion methods outperformed the baseline GMM-UBM method.