Robust state clustering using phonetic decision trees

Abstract

The widely used acoustic modeling approach of phonetic decision-tree based context clustering does not take full advantage of limited training data, and therefore fails to produce robust acoustic models. Two problems are identified: (1) all states clustered in a leaf node must share the same set of Gaussian components and mixture weights; no distinction is provided among those states; (2) rarely seen triphones in the training data might be poorly estimated and cause an adverse effect on decision-tree clustering. We propose a number of approaches to address these problems by more efficient use of training data. Specifically, (1) a two-level decision-tree approach for the first problem that ties Gaussian components and mixture weights separately, as they require different amounts of data to obtain robust estimation of their parameters; and (2) a two-stage decision-tree based clustering approach and a MAP-based approach for the second problem. Each approach gives a statistical significant reduction of the word error rate (WER) over the traditional approach. The systems combining all new approaches achieve the best performance, which reduce the WERs of the baseline systems by 14–17% and reduce the model sizes by 8–11% on the WSJ tasks.