A hybrid algorithm for estimation of the parameters of hidden Markov model based acoustic modeling of speech signals using constraint-based genetic algorithm and expectation maximization

The conventional method for estimation of the parameters of Hidden Markov Model (HMM) based acoustic modeling of speech signals uses the Expectation-Maximization (EM) algorithm. But the EM algorithm is highly sensitive to initial values of model parameters and does not guarantee convergence to a global maximum resulting in non-optimized estimation for the HMM and lower recognition accuracy. We propose a Genetic Algorithm (GA) based EM learning method (GA-EHMM) for estimation of the HMM parameters. GA explores the search space more thoroughly than that of the EM algorithm and enables the EM to escape from many local maxima. A constraint-based approach of GA has been adopted in "GA-EHMM" which directs GA towards promising regions of the search space. Instead of generating the initial GA population randomly, a variable segmentation technique is used in the HMM initialization process. "GA-EHMM" has been tested on the TIMIT [10] speech corpus. Experimental results show that "GA-EHMM" obtains better values for the likelihood function as well as higher recognition accuracy than that of the HMM model trained by the standard EM algorithm. © 2007 IEEE.