A dependence tree model of phone correlation

This paper introduces a new statistical model for acoustic observations in speech recognition. The model represents intra-utterance phone correlation and can be viewed as a probabilistic model of an utterance or a speaker. The phone correlation is modeled by a dependence tree, using the mutual information between pairs of phones as a measure of correlation. The experiments presented focus on robust tree topology design and on robust estimation for a given topology. With appropriate design algorithms, the dependence trees are shown to provide a better model for independent test data than an independent phone model