Application of structured composite source models to problems in speech processing

Abstract

An extension of the hidden Markov framework which may lead to substantial reductions in the complexity of implementing such a framework for speech modeling and recognition is proposed. This extension is suggested by the observation that speech statistics exhibit temporal structure over multiple time scales. Such temporal variation leads naturally to a special structure for the HMM (hidden Markov model). The structured composite source (SCS) is introduced as a generalization of the HMM. Theorems are developed for representing an arbitrary HMM as an SCS using techniques developed for multiple time scale analysis of weakly coupled Markov chains. Modification of the algorithms for the estimation of HMM parameters from sample data, the forward-backward and the baum-Welch algorithms, is straightforward, and results in a significant reduction in the computational complexity of the reestimation procedure. >