Markov Random Field Linear Regression

Xintian Wu, Yonghong Yan

Abstract: This paper outlines the Markov Random Field Linear Regression (MRFLR) algorithm, which combines the transformation-based adaptation and dependency-modeling technique together. The hypothesis is that the adaptation performance can be improved by explicitly modeling the correlations among acoustic parameters and applying such constraints to the transformation-matrix estimation. The correlations are modeled by Markov Random Field, and the incorporation of the correlations is under the Maximum A Posteriori framework. Experimental results show that MRFLR has significant improvement over Maximum Likelihood Linear Regression when only small amounts of adaptation data are available.