The goal of this study is to understand what makes two voices sound similar and how to quantify the degree of similarity between two voices. Previously, it was shown that the relative entropy between two voices is able to identify similar voices better than dissimilar voices. It is hypothesized that dissimilar voices differ in acoustics along multiple dimensions, and listens may apply different weights to different dimensions. In this study, weighted relative entropy was used to model the differential weighting of the different acoustic dimensions. Different weighting functions were explored, and the weighting function parameters were optimized to improve the agreement between the weighted relative entropy and the perceptual similarity scores from a listening experiment. The results showed a moderate improvement in the agreement between entropy-based and perceptual similarity scores, indicating further research is required to identify the optimal weighting function and parameters.
J Zhang (Wed,) studied this question.