Research On Speech Recognition Based On Deep Learning

With the development of computer and speech recognition technology widely used in people’s daily life,scientists paid more and more attention to the development of speech recognition technology.The traditional speech recognition technology is very dependent on the accuracy of feature extraction what mainly relying on manual skills,so its accuracy is not stable and difficult to improve.However,the technology of deep learning solved the problem.This thesis summarizes the main research results of speech recognition based on deep learning.Firstly,it analyzes the basic principle recognition of deep learning and neuron structure,elaborating the brain auditory nervous system to hand of the speech signal processing.Secondly,it compars advantages between deep learning and shallow traditional layer neural network.Deep learning uses multi hidden layer to feature learning,while shallow traditional layer neural network uses single handen layer method.So the character can be obtained better on describeing the nature of the events.Finally,it analyzes several calculation models of deep learning.Starting with the automatic encoder model and the self noise encoder model of deep learning,it proves that speech recognition based on deep learning can make the data more primitive expression compared with the traditional speech recognition,and the obtained data improves with six percent accuracy through speech recognition experiments and its results.The innovation of this thesis is that the new deep learning technology is used to the study of speech recognition,conducting the work of speech recognition through simulating the structure and working mechanism of human brain auditory nervous system.Extracting and training characteriaticss of each layer by using automatic encoder model unsupervised method.Using top-down label supervised learning method to compare samples in the interlayer.Trimming parameters,to get the best parameter sets,so that it can improve the efficiency and accuracy of speech signal feature training.