Research Of Speech Recognition Algorithm Based On Neural Network

Neural network or artificial neural network is a kind of intelligent nonlinear processing system.The main difference between it and the traditional pattern recognition algorithm is that its information processing stage mimics the information processing of human brain nervous system.It mainly enhances its adaptability to information intelligence processing through simple distributed parallel.In the process of processing,it has the functions of storage,association and retrieval,etc.Especially in speech recognition and other sensitive issues have strong processing power.The new neural network recognition system is more dominant and scientific,and can be used to store voice information divergently in a similar way with human's brain.This thesis mainly studies neural network based speech recognition system.Firstly,the principle of speech recognition and neural network is introduced,and a simple speech recognition model based on BP neural network is constructed.A large number of experiments are performed on the performance and recognition effect of the model.After analyzing the preprocessing process of speech signal in detail and summarizing the shortcomings of the traditional feature parameter extraction method,this paper proposes an optimization method-the hybrid MFCC extraction method.According to the deficiency of the traditional BP neural network algorithm,it proposes to increase the momentum.The combination of factors and optimization using adaptive learning rates.The main work and contributions of the thesis are summarized as follows:This paper mainly studies the speech recognition system based on neural network.Firstly,the principle of speech recognition and neural network is introduced.Then,the preprocessing process of speech signal is analyzed in detail,and the shortcomings of the parameter extraction method with traditional feature are summarized.On this basis,the optimization method of this paper,the hybrid MFCC extraction method,is explained.Then the application of BP neural network in speech recognition and the advantages and disadvantages of traditional BP neural network algorithm are stated.In view of the shortcomings of the traditional BP neural network algorithm(that is,the local minimum is easy to appear and the convergence speed is not fast enough),this paper proposes an optimization method combining the increasing momentum factor and the use of adaptive learning rates.At last,a complete speech recognition system is constructed,and many simulation experiments are carried out,and the results are given.It provides a new idea for the improvement and optimization of speech recognition.The innovations of this paper are as follows:(1)In view of the different recognition effect of the three characteristic parameters of IMFCC / Mid MFCC MFCC in different speech band,the algorithm used in this paper is a hybrid algorithm which combines the three parameters,and the filter banks are used to construct the characteristic parameters.Compared with the traditional LPCC parameters,the traditional MFCC parameters such as mixed parameters of MFCC and MFCC have better recognition performances.The recognition accuracy of each speech band is improved effectively.(2)Local minimization and slow convergence are the most common disadvantages of traditional BP neural network algorithm.In this paper,the methods of increasing momentum factors and using adaptive learning rates are adopted to optimize BP algorithm.Experiments show that compared with the traditional BP algorithm,the improved algorithm can effectively prevent BP network from falling into local minimum and improve the slow convergence speed of neural network.This thesis mainly studies the speech recognition system based on BP neural network algorithm.The traditional MFCC algorithm is optimized,and the application of BP neural network algorithm in speech recognition is improved.Use isolated numbers 0-9 as identification objects,and the improved algorithm is verified by experiment and simulation.