Single Channel Speech Enhancement Based On Generative Adversarial Networks

Speech signal is one of the important means of information transmission in people's production and life.In the process of information transmission,speech signals will inevitably be polluted by various noises,resulting in varying degrees of information loss.In order to restore the speech information polluted by noise as much as possible,reduce the loss of speech information in the transmission process,and improve the intelligibility and quality of speech,speech enhancement technology has been proposed and widely used in military communications,teleconference and other scenarios.As the first step of speech signal processing,speech enhancement effect will directly affect the communication quality.However,existing speech enhancement technologies still have some problems,such as poor enhancement effect and poor model generalization ability in low SNR environment.The main goal of this paper is to propose a speech enhancement model based on multi-scale generative adversarial networks.This model still has strong noise reduction ability and generalization ability in low SNR environment.Different from the original speech enhancement model based on generative adversarial networks,this model generates speech signals of multiple dimensions in the generator and discriminates them in multiple sub-discriminators.It can enhance speech signal from multiple dimensions to enhance the effect of speech enhancement in low SNR environment.The main research work and innovation points of this paper are as follows:(1)The principle of speech enhancement algorithm based on generative adversarial networks is introduced in detail.Simulation experiments are carried out on the spectrum subtraction method,the improved logarithmic spectrum amplitude estimation algorithm based on MMSE and the speech enhancement method based on the generative adversarial networks,the existing problems and advantages of this algorithm are analyzed.(2)A new structure of multi-scale generator and discriminator system is proposed to solve such problems as unstable confrontational training of the original speech enhancement algorithm based on generative adversarial networks and poor noise reduction performance in a low SNR environment.The algorithm can enhance the speech from different dimensions and improve the overall speech enhancement effect.The experimental data show that this method has better performance than the original algorithm and the traditional method,especially in the case of low SNR.(3)In order to prevent the insufficient use of speech information and reduce the loss of information in the network,we added the down-sampling hopping connection to the model to enrich the voice information,and obtained the optimal hopping connection combination through experimental simulation.In order to improve the enhancement effect of the network's final output speech,we increase the weight ratio of the original dimension to generate the speech signal,so that the network can reduce the speech noise from multiple dimensions and focus on the generation of the original dimension to enhance the speech.(4)We took the improved logarithmic spectral amplitude estimation algorithm based on MMSE optimal as the pre-processing step,and combined it with the speech enhancement model based on multi-scale generative adversarial networks to form the joint speech enhancement algorithm.The pre-processing step can reduce the interference of noise signal to multi-scale generation antagonistic network,improve the intelligibility and SNR of speech signal,and provide more effective features for multi-scale generative adversarial networks.