A Text-Independent Speaker Verification System Based on Cross Entropy

This paper presents a method based on information theory to estimate the distortion between the enrolled speaker's model and the test utterance in speaker verification system. It uses the cross entropy (CE) to compute the distance between two parametric models (such as GMMs). Different from the traditional average log-likelihood method, it considers the symmetry between the test utterance and the referenced model. In the verification phase, the Zt - norm is used to compensate the session variability. Experiment results based on the TIMIT database show that the proposed method can efficiently reduce error rates over the standard log-likelihood scoring.