Application Of Deep Learning In Speech Enhancement

In recent years,with the rapid development of computer and digital signal processing technology,speech communication has gradually become an important way in human-computer interaction.However,all kinds of interference noise can cause problems of delay or wrong recognition of human-machine interaction.Therefore,it has practical significance to solve noise interference and enhance speech quality.The traditional enhancement method will remain part of the noise,especially the music noise.In addition,the traditional method has made some assumptions about the relationship and characteristics of signals,which is not in line with the situation of actual environment.In this paper,firstly,the basic knowledge and related theories of speech enhancement technology and deep learning are thoroughly studied,including the preprocessing of speech signal,feature extraction and training of deep learning model.Secondly,the traditional speech enhancement framework based on deep learning is given,which takes the features of the noisy speech signal as the input of the deep neural network,the characteristic of the pure speech signal as the output of the network,and uses the powerful feature extraction and mapping ability of the deep neural network to reproduce the pure speech signal.Secondly,the spectral subtraction is a more classic method in traditional speech enhancement algorithms.Its computation is small and easy to be realized,but its sound quality is not ideal,and there is a lot of music noise in the enhanced speech signal.In this paper,the traditional spectral subtraction and deep network are used to reduce the noise.By spectral subtraction,most of the noise doped in the speech signals can be removed easily.The residual noise,especially the music noise,is filtered out by the self coding of the stack.In the experiment,4 noises are randomly selected from the database to generate 6kinds of noisy speech signals with different signal-to-noise ratios.The experimental results show that the two methods can significantly reduce the noise in the speech signal,improve the speech perception,and have a certain generalization ability and adaptability for the unknown noisy environment.