Statistical approaches to acoustic modeling in speech recognition based on discrete HMM

Abstract

Recently, hidden Markov model (HMM) has become the predominant approach to speech recognition. The performance of an HMM-based recognizer depends on the acoustic modeling techniques adopted to represent speech characteristics. In this dissertation work, we present various methods to improve acoustic modeling capability in a discrete HMM-based speech recognition system. In order to evaluate the recognition performance for each method, we established two kinds of baseline speech recognition systems. One is an isolated word recognition system in which the vocabulary consists of 75 words, and the other is a continuous speech recognition system with a vocabulary of 102 words. First, we study various methods to estimate robust output probability distributions (PD``s) with insufficient training data. We take the approach to interpolate an output PD with some general PD``s which are able to strengthen the robustness with respect to unseen data. For this approach, we propose a statistically reliable deleted interpolation (DI) method which improves the DI technique conventionally used for robust PD interpolation. On the other hand, in order to enhance the robustness of a PD against unseen data, we study two approaches: smoothing and tying. We introduce a new algorithm to smooth a PD where the smoothing matrix is estimated by following the strategy of cross-validation as in DI. Furthermore, the smoothing method called the deleted smoothing (DS) method resembles the probabilistic mapping used for speaker adaptation. As for tying, we derive a number of state clusters based on a clustering tree which achieves a good compromise between robustness and detail of the tied PD``s in specifying speech characteristics. In addition to providing an efficient method for constructing the clustering tree, we propose a measure which accounts for the variability of estimated PD``s under various situations. In isolated word recognition experiments, the proposed methods reduce the error rate...