Decision tree-based simultaneous clustering of phonetic contexts, dimensions, and state positions for acoustic modeling

In recent years, context-dependent hidden Markov model, typically triphones and continuous density HMMs are often used. The use of triphones results in too many free-parameters in a system, hence it is difficult to estimate the model which is statistically reliable. Therefore, various parameter clustering techniques have been proposed. The use of Phonetic Decision Trees (P-DT) based state-tying technique is a good solution to this problem. However, state-tying technique cannot construct proper context-dependent sharing structure and cannot assign proper number of free-parameter for each dimension. In this paper, Phonetic and Dimensional Decision Trees (PD-DT) is proposed by introducing the MDL-based dimensional-split technique into P-DT. Furthermore, by incorpolating questions about state positions into PD-DT, Phonetic, Dimensional and State positional Decision Trees (PDS-DT) is defined. In speaker-independent continuous speech recognition experiments, proposed technique achieved about 13%–15% error reduction over P-DT based state-tying technique.