Sixth International Conference on Spoken Language Processing
A good similarity measure of random variables is crucial in many applications. The choice of distance measure directly affects quality of system design. In this paper, we present a new measure of the similarity between two random variables. The discussion here emphasizes on the case of normal distribution. Based on this Gaussian Similarity Analysis (GSA), we propose an algorithm in speaker adaptation of covariance. It is different from the traditional algorithms, which mainly focus on the adaptation of mean vector of state observation probability density. A binary decision tree is constructed offline with the similarity measure and the adaptation procedure is data-driven. It can be shown from the experiments that we can get a significant further improvement over the mean vectors adaptation.
Bibliographic reference. Wu, Ji / Wang, Zuoying (2000): "Gaussian similarity analysis and its application in speaker adaptation", In ICSLP-2000, vol.4, 370-373.