|
As the computer has become more and more portable, the environment of that has been more complicated, human desire to communicate with computer in a portable, natural, effective way without the constraints of keyboard. In this situation, speech recognition technology has developed, which controls the computer to implement the order according to the speech. It can bring the human much convenience. At present, it has been applied extensively in human life. Using modern methods to study the speech recognition technology, it can effectively help us produce, storage and retrieval of speech signal, and it also promotes social development with great significance.The speech recognition system consists of three main parts, namely, preprocessing, feature extraction and recognition unit. Among them, selecting the better characteristic parameters of the speech and the recognition unit is particularly important.MFCC parameters which are based on the characteristics of human hearing’s as features of speech signals have two shortcomings, first of all, they can not reflect the dynamic nature of the speech signals, secondly, they cannot describe the characteristics of high-frequency speech signals accurately, to solve these problems, this paper introduces△MFCC and IMFCC parameters, combined with MFCC to compose the mixed parameters MFCC+△MFCC+IMFCC as input of the recognition network. In addition, this paper used a new type of recurrent neural network-ESN as it’s recognition network, And use PSO algorithm to optimize the parameters of the dynamic reserve pool and the output weight matrix. The results of the simulation for speech recognition of isolated words which is from non-specific shows that the mixed parameters and the model we select can improve the recognition rate effectively.In this paper, the specific recognition algorithms simulation is conducted in MATLAB software, firstly, this paper preprocessing the speech signal; secondly, this paper have done two groups of simulation experiments based on LPCC, MFCC and the mixed characteristic parameters respectively, The first group of the experiments proved the mixed characteristic parameters which is based on MFCC can better reflect the characteristics of the speech signal, The second group of the experiments proved the mixed characteristic parameters which is based on MFCC has better noise immunity; At last, this paper have done the simulation experiments based on ESN and the improved ESN, the result shows that the improved ESN can improve the recognition rate effectively.
|