Robust decision tree state tying for continuous speech recognition

Abstract

Methods of improving the robustness and accuracy of acoustic modeling using decision tree based state tying are described. A new two-level segmental clustering approach is devised which combines the decision tree based state tying with agglomerative clustering of rare acoustic phonetic events. In addition, a unified maximum likelihood framework for incorporating both phonetic and nonphonetic features in decision tree based state tying is presented. In contrast to other heuristic data separation methods, which often lead to training data depletion, a tagging scheme is used to attach various features of interest and the selection of these features in the decision tree is data driven. Finally, two methods of using multiple-mixture parameterization to improve the quality of the evaluation function in decision tree state tying are described. One method is based on the approach of k-means fitting and the other method is based on a novel use of a local multilevel optimal subtree. Both methods provide more accurate likelihood evaluation in decision tree clustering and are consistent with the structure of the decision tree. Experimental results on Wall Street Journal corpora demonstrate that the proposed approaches lead to a significant improvement in model quality and recognition performance.