Speaker verification with ensemble classifiers based on linear speech transforms
Jesper Østergaard Olsen · 1998
For most classifier architectures realistic training schemes only allow classifiers corresponding to local optima of the training criteria to be constructed. One way of dealing with this prob-lem is to work with classifier ensembles: multiple classifiers are trained for the same classification problem and combined into one “super ” classifier. The problem addressed in this paper is text prompted speaker verification by means of phoneme dependent Radial Basis Function networks trained by gradient descent error minimisation. In this context ensemble techniques are introduced by combining different classifiers that classify feature vectors, which have been pre-processed using different linear transforms. Four different types of linear transforms are studied: the Fisher transform, the LDA transform, the PCA transform and the co-sine transform. The verification system is evaluated on the Gan-dalf database, where the equal error rate is reduced from 3.6 % to 3.2 % when ensemble techniques are introduced. 1.