HMM training by a hybrid of Chaos Optimization and Baum-Welch algorithms for discrete speech recognition

HMM has high power to describe complex phenomena. The Baum-Welch (BW) algorithm is very popular estimation method that use for estimating HMM model parameters but it start with an initial guess and finally converge to a local optimum in practice. Chaos often exists in nonlinear systems. It has many good properties such as ergodicity, stochastic properties, regularity and high sensitivity to initial states. In this paper by use of these properties of chaos, an effective hybrid CHAOS-BW optimization method is proposed that uses the Chaos Optimization algorithm to optimize the initial values of Baum Welch algorithm. This algorithm not only overcomes the shortcoming of becoming trapped in local optimum of the BW algorithm, but is also fast and requires less storage than other hybrid optimization algorithms such as GABW, PSOBW and GAPSOBW. Experimental results on Persian digit dataset show that the propose method has both qualities of global search as well as rapid convergence. Comparison with several other more conventional approaches also reveals superior performance of the proposed model.