Research On Speech Enhancement Method Based On Generative Adversarial Networks

The purpose of speech enhancement is to eliminate background noise and environmental interference in speech signals,to improve the quality of polluted speech,and to extract pure and unpolluted speech as far as possible.It plays an important role in speech recognition system and communication system.However,due to the diversity of the environment,the instability and randomness of noise,and the basic impossibility of obtaining all sample data,these methods are difficult to achieve better results in practical application.In addition,the traditional enhancement methods will make some assumptions about the distribution of speech signals.The inaccurate assumptions will make the difference between the enhanced speech and the pure speech signal larger,and the distortion of the speech is relatively high.As a result,the intelligibility of the speech is not high and the quality of the speech decreases.Based on these problems,this paper research speech enhancement model based on generative adversarial network.Generative adversarial network belongs to generative model,and zero-sum game is used to train generators and discriminators.Its generator does not need to make any assumptions about the distribution of data.The greatest advantage is that it can learn the real data under any distribution and generate data with similar distribution.In addition,most of the generated adversarial networks are constructed with neural network structure,which not only improves the generalization ability of the model,but also accelerates the training speed of the network countermeasure,and has great advantages.The main work of this paper includes the following two aspects:(1)The selection of the loss function of the Generative Adversarial Network is closely related to the performance of the network model.By analyzing the problems faced by the loss function of the original generated countermeasure network in the process of model training,the least squares loss function is selected.Aiming at the problem of low quality speech signal reconstructed by generator,an improved least squares loss function is proposed.The L1 norm is introduced into the loss function of the generator,and the weight coefficients of the influence of L1 regularization are controlled by hyper-parameters,and the optimal values are obtained through experiments.According to the characteristics of speech signal,the network structure of discriminator and generator based on speech enhancement is built on the basis of improvement,and the performance of the model is trained and tested.Compared with the traditional speech enhancement algorithm,it is concluded that the speech enhancement effect based on the Generative Adversarial Network is better than the traditional speech enhancement algorithm.(2)On the basis of deep convolution Generative Adversarial Network,this paper improves the network structure of generator and discriminator,aiming at the slow convergence speed and the disappearance of gradient in the training process.Based on conditional Generative Adversarial Network,the high-dimensional abstract features extracted by the discriminator are used as conditional information of the generator,which is input to the generator together with Gauss noise,and the conditional Generative Adversarial Network model is constructed.Finally,by comparing and analyzing the speech enhancement algorithm based on the Generative Adversarial Network before the improvement,it is found that the generated antagonism network can obtain better enhancement effect,the enhanced speech quality has been improved,and the stability and generalization ability of the algorithm have also been increased.