SVM-based speaker verification for coded and uncoded speech

This paper describes experiments with speaker verification using support vector machines (SVMs). Verification from coded and uncoded speech is analyzed, both in matched and mismatched conditions. A hybrid SVM-GMM approach is used, in which SVM classifiers with Kullback-Leibler kernel make verification decisions based on the mean values of Gaussian mixtures. The most common narrowband codecs are used, such as G.711, G.729, G.723.1, GSM 06.10, GSM 06.60, and Speex. The Equal Error Rate (EER) is presented for various numbers of Gaussian components, and for various testing conditions. Possible reasons for the non-uniform performance degradation in the case of codec mismatch are discussed. Selected ROC curves are presented. The results are compared with a similar investigation of a close-set speaker classification.