Research On Speech Enhancement Method Based On Deep Neural Network

Speech enhancement plays an important role in the field of digital signal processing.It can improve the quality of damaged speech and enhance its intelligibility.It has been widely used in the smart home.With the popularity of deep learning in the computer field,network model training based on deep learning has become the mainstream method of speech enhancement.At present,it is mainly based on deep learning to build a network model,and combined with a large number of data operations,in the model to learn the function from the noisy features to enhance the target,in order to solve the task of speech enhancement.However,the scheme still has some shortcomings,such as low quality of speech enhancement,slow model training speed and low score of model evaluation index.In order to solve the above problems,a series of research on speech signal enhancement based on deep neural network is carried out in this paper.Firstly,the speech enhancement model based on deep neural network is studied,and the loss function used in network training is analyzed.Because the correlation between speech frames is not fully utilized,the enhancement effect of the model is not good enough.Therefore,on this basis,an optimized loss function is introduced,combined with the deep neural network training model to make full use of the correlation between adjacent frames of speech signal.Simulation results show that the speech enhancement effect of the proposed training scheme is significantly better than that of the original training scheme and the traditional model training scheme,which greatly enhances the speech quality and intelligibility.Then,this paper proposes to use the U-Net model in the field of image to build an improved model for training speech enhancement by using the end-to-end characteristics of U-Net model and residual network.The innovation is to introduce the residual unit into the codec block of the model,that is,to apply the cross layer connection of residual network structure and fitting residual term to the model training.This method is more conducive to recovering the detailed feature information of the target speech,enhancing the stability of model training,improving the feature extraction ability and training efficiency of the model.The improved residual-U-Net network model can achieve better speech enhancement effect.Simulation results show that:compared with other existing speech enhancement methods,the proposed residual-U-Net algorithm can improve speech quality,reduce speech distortion,and has a certain noise reduction effect.To sum up,this paper mainly explores the single channel speech enhancement algorithm based on deep neural network.Experimental results show that the proposed algorithm can further improve the speech quality and intelligibility compared with the traditional methods.Finally,it summarizes the content of this paper and puts forward the future research direction and trend..