A discriminative decision tree learning approach to acoustic modeling

The decision tree is a popular method to accomplish tying of the states of a set context dependent phone HMMs for efficient and effective training of the large acoustic models. A likelihood-based impurity function is commonly adopted. It is well known that maximizing likelihood does not result in the maximal separation between the distributions in the leaves of the tree. To improve robustness, a discriminative decision tree learning approach is proposed. It embeds the MCE-GPD formulation in defining the impurity function so that the discriminative information could be taken into account while optimizing the tree. We compare the proposed approach with the conventional tree building using a Mandarin syllable recognition task. Our preliminary results show that the separation between the divided subspaces in the tree nodes is clearly enhanced although there is a slight performance reduction.