Speech Recognition Research Based On Neural Network And Acoustic Characteristics Of Hybrid Structures

In recent years,computer technology has a rapid develop,and neural network has also ushered in a new round of upsurge,based on neural network hybrid structure model and acoustic feature extraction have made remarkable achievements.Since the first time in2011,Microsoft Research Institute has used deep neural network(DNN)in the large-scale speech recognition task to obtain significant results to enhance the neural network in the field of speech recognition gradually more and more attention from all walks of life.The purpose of speech recognition technology is to let the machine can understand people say,due to computer technology has the rapid development,people once again on the speech recognition technology hope.Firstly,this paper introduces some generalizations of speech recognition technology,and then discusses the feature extraction of speech recognition.In view of the traditional static characteristic parameters(LPCC,MFCC)and dynamic characteristic parameters(LPCC,MFCC(LPMFCC),the corresponding theory and algorithm are analyzed.At the same time,the progress of the proposed neural network is introduced.Secondly,it discusses the acoustic modeling of neural network and its role in speech recognition.It summarizes the development process of acoustic modeling of neural network,focusing on the details of Restricted Boltzmann Machine(RBM),commonly used learning Algorithms,and evaluation methods are described systematically.The speech recognition experiment based on backward propagation algorithm(Back-Propagation)is studied by using MFCC and(LPMFCC,MFCC)mixed parameters in BP neural network and DBN neural network acoustic model.The results show that the linear prediction plume The characteristics of the cepstrumn coefficient can be better than the Meier cepstrum coefficient,and the BP neural network in the recognition system can be optimized and shortened The training time,improve the recognition performance.RBM model is easy to learn in depth learning algorithm,this model of the algorithm to overcome the direct multi-layer network training efficiency.In this paper,the Deep Belief Nets(DBN)model is built by RBM stack in speech recognition experiment.With the deep belief network,the continuous multi-frame speech feature can be used together,because the multi-layer structure of the human brain can be simulated,so the information feature extraction can be carried out step by step.In the DBN acoustic model,we use the time-adjusted MFCC and(LPMFCC,MFCC)mixed parameters as input data.The experimental process is based on RBM setting rules to optimize the network model,enhance the model learning effect,and with the traditional BP neural network The model is compared and found to achieve a better recognition effect.