Discriminant NAP for SVM speaker recognition

Nuisance Attribute Projection (NAP) provides an effective method of removing the unwanted session variability in a Support Vector Machine (SVM) based speaker recognition system by removing the principal components of this variability. There is no guarantee with the methods proposed, however, that desired speaker variability is retained. This paper investigates the possibility of training NAP discriminatively to remove session variability while maintaining desirable speaker variability through an approach which is a variation on Scatter Difference Analysis (SDA). Experiments on NIST SRE tasks with a GMM mean supervector SVM system demonstrate a modest improvement by using SDA for NAP training by adding some speaker scatter.