Research And Implementation Of Deep Neural Network Model In Speech Recognition

The communication between humans and computers is inseparable from speech recognition.This kind of human-computer interaction not only has higher efficiency,but also can carry out complex emotional expression and greatly enhance users experience in human-computer interaction.In the past few years,the technological revolution brought about by artificial intelligence and deep learning has affected our lives extensively.The explosion of technology cannot be separated from the improvement of data quality and quantity,but also the progress of model algorithms.This article starts from the basic theory of speech recognition technology and mainly introduces the preprocessing,feature processing,acoustic model and language model of speech signal.The most widely used shallow acoustic model HMM and n-gram language models in acoustic models are analyzed.In order to deal with large-scale speech corpus,the shortcomings of modeling and recognition are insufficient.Derived the concept of deep neural networks.Compared with traditional GMM-HMM,DNN-HMM has a significant performance improvement in all aspects.It has the characteristics of being able to handle large-scale data and of high recognition rate,and is widely used in the industry.This paper starts from the composition of the deep neural network model,describes in detail the working mechanism of neurons,the basic component in deep neural network,and the back propagation algorithm(BP algorithm),and studies the model optimization and implementation.Different loss functions and activation functions have an greater impact on the model performance.For the combination of the traditional mean square error loss function and sigmoid activation function,the gradient declines slowly during training.This results in slower convergence of the model,and a combination of cross-entropy loss function and ReLU activation function can effectively eliminate this effect.By thoroughly studying the two methods of preventing overfitting and regularization,dropout can balance different neural networks in training and avoid overfitting by avoiding the mutual adaptation of hidden layer units.The idea of using dropout is proposed to reduce the neurons in the hidden layer of the network model to reduce parameters and prevent overfitting in the model construction.The feasibility of this idea is verified in practical training.