Novel Lookahead Decision Tree State Tying for Acoustic Modeling

This paper presents two new lookahead methods of constructing phonetic decision trees (PDTs) for acoustic model state tying, a constrained method and a stochastic method. The constrained lookahead method searches for optimal phonetic questions among pre-selected question sets, and reduces contributions of deeper decedents as a function of their levels in the tree. The stochastic full lookahead method uses subtree size instead of likelihood gain as a judgment in selecting a phonetic question for a node split, in order to find a compact tree that is consistent with training data. Since the computational cost of exhaustive lookahead is prohibitively high, a stochastic subtree generation method is used to explore most promising question at each node. We also propose using a phone-state dependent threshold instead of a fixed threshold of likelihood gain to decide if a node split should continue or not. Furthermore, we use a fast confusion network (CN) algorithm to combine recognition hypotheses produced by using acoustic models from different PDT training methods. Experimental results show that the proposed lookahead methods consistently decrease model size, and the integration of recognition hypotheses consistently improves recognition accuracy.