Maximum mutual information based reduction strategies for cross-correlation based joint distributional modeling

In maximum-likelihood based speech recognition systems, it is important to accurately estimate the joint distribution of feature vectors given a particular acoustic model. In previous work, we showed we can boost the accuracy in this task by modeling the joint distribution of time-localized feature vectors along with the statistics relating those feature vectors to their surrounding context. In this work, we evaluate information preserving reduction strategies for those statistics. We claim that those statistics corresponding to spectro-temporal loci in speech with relatively large mutual information are most useful in estimating the information contained in the feature-vector joint distribution. Furthermore, we claim that such statistics are most likely to generalize. Using an EM algorithm to compute the mutual information between pairs of points in the time-frequency grid, we verify these hypotheses using both overlap plots and speech recognition word error results