The Research Of Speech Recognition Base On GA-ACO Algorithm And BP Neural Networks

As the most important interpersonal communication tool, voice has played a very important role in the constitution of human's artificial intelligence. With the rapid progress and development of computer technology and communication technology, it has a good prospect in various industries, so it attracts people's attention attentively. In the past, the study of voice was traditionally based on linear theory, we often use the method of Dynamic Time Warping(DTW) and Hidden Markov Model(HMM).However speech recognition process is not a simple linear process, and its nonlinear makes the defects of identification method using linear theory unfold slowly. In recent years, there were more and more researches of the neural network and the nonlinear theory of it was understood better by people. Being used in speech recognition, it can get better results, therefore, it has also become the focus of research today.This paper expounds the whole process of speech recognition: preprocessing, feature extraction patterns training and match. Preprocessing includes filter-sampling, pre-emphasis, frame add window and endpoint detection; Feature extraction states three basic methods: Linear Prediction Coefficients(LPC), Linear Prediction Cepstrum Coefficient(LPCC) and MEL Frequency Cepstrum Coefficient(MFCC); Patterns training and match expounds the Dynamic Time Warping(DTW) and Hidden Markov Model(HMM) and Artificial Neural Networks(ANN). This paper mainly focuses on the further study of the principle of feed-forward neural network(BP) and its application in speech recognition and analyzes its shortcomings, then putting forward the combination of Genetic Algorithm(GA) and Ant Colony Optimization(ACO).The optimized BP neural network overcomes some disadvantages such as slow convergence speed, local minimum value, unfavorable dynamic performance, the restriction of study precision. Thus getting a new way to implement speech recognition, and in this paper, after the simulation experiment using this method, its feasibility was validated.For the unspecified person and separate digital identification, considering the shortcoming of BP neural network, this paper uses GA-ACO fusion algorithm for its training according to the basic methods of speech recognition to form a new speech recognition method and the simulate the different results using the new method and the traditional BP neural network method for speech recognition, then discuss the influence of identifying the accuracy with different characteristic parameters, the training set, the number of hidden layer neurons.The research results show that the proposed neural network based on GA-ACO algorithm identification method can learn network weights quickly and eliminate the difficulty that traditional BP algorithm can easily fall into local extremum and how to select the initial weight value to quickly reach setting precision. From all aspects it can be better than the traditional BP algorithm, it has a higher recognition accuracy and better application.