Speech recognition is a hot issue in the field of speech signal processing, and always is difficult, especially for the independent person in the noisy environment. The training unit is a very important part in speech recognition and it determines the entire properties of the system.As a statistical model in description of speech signal, HMM is widely used in signal processing especially in speech processing. This article illuminates the principle of HMM and the solution of its three questions. Training problem is the most difficult and important problem of the three problems. The essence of classical training algorithm Baum Welch is grads descending method, which may reach a locally minimal solution in the training process. So, the selection of initial value of the model is important, and good initial value could avoid locally minimal solution.An important characteristic of genetic algorithm is global search, so we can get globally optimal solution or at least sub-optimal solution. Selection, crossover and mutation are three main operators of genetic algorithm, and individual is the object of operation. They comprised of the whole process of inheritance and make genetic algorithm have the eminent trait which other classical methods don\’t have. Because of the great effect of initial value of B for the HMM, this paper applied genetic algorithm to the optimization of initial value of B in Baum Welch and proposed a hybrid algorithm combined classical algorithm with genetic algorithm.This paper first encoded the initial value of B, then designed fitness function, crossover operator, mutation operator and selected the control parameter, and finally simulated the above genetic algorithm, so the optimal initial value of B was obtained. Then the model was trained by Baum Welch, and at last was recognized by Viterbi. In the experiment, because the performance of genetic algorithm is greatly related to the selection of its control parameter, the optimal performance of genetic algorithm always needs optimal parameter. The experiment carried out a search on the two most important parameters of P_c and P_m. Eventually, the optimal model was obtained, and therecognition rate of the entire speech recognition system was elevated.

Related posts:

1. The Algorithm Research of Speech Recognition Based on VQ and DTW
2. The Research of Feature Extraction Algorithm for Speech Recognition and the Realization
3. A Research of the Chinese Speech Recognition Based on the GA-BP Model
4. Research on Improved Performance of Ant Colony Algorithm by Genetic Algorithm
5. Speech Recognition System Based on TMS320C5409