Study On The Speech Model Optimization Based On LSTM Neural Network

With the rapid development of machine intelligence,speech intelligent recognition technology,as a means to improve the efficiency of human-computer interaction,has become an inseparable part of the field of machine intelligent technology.Its application is increasing,and it is widely accepted and widely used by the public.However,in the development of the speech recognition model,a large number of computing resources and training time are needed,which has become a limiting factor for the further improvement and development of speech recognition technology.Therefore,it is of great theoretical and practical significance to quickly train models and reduce the required computing resources in the development stage of speech recognition models.This paper mainly studies the main factors in the development of the speech recognition model of memory recurrent neural network based on long and short memory,which causes the slow training speed of the long and short memory recurrent neural network,and how to speed up the training speed of the network and reduce the computing resources needed to reduce the network.The main research work of this paper is as follows:1.This paper puts forward the method of adding classifier to the basic recurrent neural network output layer to decompose the calculation process of the output layer,and improve the complexity of the operation process of the recurrent neural network output layer,so as to reduce the complexity of the network output layer.The experimental results show that the method of adding the classifier's output layer structure to the output layer of the network can reduce the computational complexity of the output layer of the network.2.The threshold activation values in the long and short term memory circulation neural network are sparse,and there is a certain linear relationship between some threshold activation values.In this paper,a method of linear representation of negative gate activation is proposed to reduce the number of gate activation,reduce the computational complexity of network gate and improve the efficiency of network learning.The experimental results show that this methodimproves the efficiency of network training on the premise of ensuring certain identification precision.3.In this paper,the length of the improved neural network and the neural network(CNN)memory cycle,FSMN neural network are analyzed in the experiment,the experimental results show that based on the improved memory cycle when the length of the speech recognition model of neural network recognition performance is better.