A Constraint-Based Evolutionary Learning Approach to the Expectation Maximization for Optimal Estimation of the Hidden Markov Model for Speech Signal Modeling

This paper attempts to overcome the tendency ofthe expectation–maximization (EM) algorithm to locate a localrather than global maximum when applied to estimate the hiddenMarkov model (HMM) parameters in speech signal modeling.We propose a hybrid algorithm for estimation of the HMM inautomatic speech recognition (ASR) using a constraint-based evolutionaryalgorithm (EA) and EM, the CEL-EM. The noveltyof our hybrid algorithm (CEL-EM) is that it is applicable forestimation of the constraint-based models with many constraintsand large numbers of parameters (which use EM) like HMM. Twoconstraint-based versions of the CEL-EM with different fusionstrategies have been proposed using a constraint-based EA andthe EM for better estimation of HMM in ASR. The first one usesa traditional constraint-handling mechanism of EA. The otherversion transforms a constrained optimization problem into an unconstrainedproblem using Lagrange multipliers. Fusion strategiesfor the CEL-EM use a staged-fusion approach where EM has beenplugged with the EA periodically after the execution of EA for aspecific period of time to maintain the global sampling capabilitiesof EA in the hybrid algorithm. A variable initialization approach(VIA) has been proposed using a variable segmentation to providea better initialization for EA in the CEL-EM. Experimental resultson the TIMIT speech corpus show that CEL-EM obtains higherrecognition accuracies than the traditional EM algorithm as wellas a top-standard EM (VIA-EM, constructed by applying the VIAto EM).