This paper investigates the convergence between SVM-based and distance-based classifiers in speaker recognition. It focuses on approaches in speaker recognition where a speech utterance is represented with a fixed dimension vector. We study various preprocessings to apply on the vectors before classification, and various choices of negative samples to train SVM. We prove that in one specific configuration, the SVM-based classifier and the distance-based classifier are strictly equivalent. Experiments on NIST2006 database, within the Anchor Models framework, show that this specific configuration gets very good performance.
Bibliographic reference. Charlet, Delphine / Zhao, Xianyu / Dong, Yuan (2008): "Convergence between SVM-based and distance-based paradigms for speaker recognition", In INTERSPEECH-2008, 1389-1392.