Research On Speech Enhancement Method Based On Generative Adversarial Networks

How to extract approximate pure speech in noisy environment has always been a key issue in the research of speech enhancement technology in speech signal processing.In the traditional speech enhancement methods,it is usually necessary to estimate the energy information of the noise signal first and assume the distribution of the speech signal.But in reality,noise signals tend to be time-varying and random.These assumptions often have their own problems,resulting in the effect of speech enhancement is not ideal.Therefore,it is of great significance to study a different speech enhancement method which is different from the traditional method for speech signal processing.Along with the upsurge of learning artificial intelligence(AI),all kinds of deep learning models emerge as the times require.Generative Adversarial Networks(GAN)as a deep learning model has become a hot research direction in recent years.GAN is made up of generative model and discriminative model.Both are learning and training by competing and confronting each other.The goal of GAN is to generate new data.GAN has been successfully applied to image processing and has achieved good results.But the use of GAN in speech enhancement is still in the start stage.This paper is using GAN to study how to make speech enhancement.The research content and research results of this paper are as follows:1.A speech enhancement method based on least squares GAN is proposed.This method is an end to end speech enhancement method in time domain.It does not need to estimate the energy information of the noise signal,and does not make any assumptions about the distribution of the speech signal.By comparing the speech enhancement with the traditional speech enhancement methods in the time domain,the experimental results show that the proposed method has less noise than the traditional speech enhancement method,and the enhanced speech waveform is closer to the pure speech waveform.2.A speech enhancement method based on least squares GAN in frequency domain is proposed in the case of training the speech enhancement method with large amount of data and slow speed in the time domain.The experimental results show,in this method,the training data is reduced by 50% and the speed is increased by about 29.1% compared with the least squares GAN speech enhancement method in time domain.And it is still superior to the traditional speech enhancement methods.