MDL-based context-dependent subword modeling for speech recognition

Context-dependent phone units, such as triphones, have recently come to be used to model subword units in speech recognition systems that are based on the use of hidden Markov models (HMMs). While most such systems employ clustering of the HMM parameters (e.g., subword clustering and state clustering) to control the HMM size, so as to avoid poor recognition accuracy due to a lack of training data, none of them provide any effective criteria for determining the optimal number of clusters. This paper proposes a method in which state clustering is accomplished by way of phonetic decision trees and in which the minimum description length (MDL) criterion is used to optimize the number of clusters. Large-vocabulary Japanese-language recognition experiments show that this method achieves higher accuracy than the maximum-likelihood approach.