Feature evaluation and selection for an on-line, adaptive speaker verification system

This paper describes two feature selection algorithms applied to different speech segments which are used in an on-line, adaptive speaker verification system. The Information Theoretic approach is used to reduce the redundancies among the features that are originally present in the feature pool. Next, the Between-to-Within variance ratio (BW ratio) feature ordering algorithm is applied for obtaining the optimum combination of interspeaker separability and intraspeaker variability. A preliminary set of experimental results are presented.