Research On Speech Enhancement Methods Based On Generative Adversarial Networks

Speech signals are susceptible to interference from various environmental noises during transmission,which affects the effectiveness of tasks such as speech communication and speech recognition.Speech enhancement techniques can reduce the impact of noise in speech signals and improve the quality and clarity of speech signals.Based on traditional signal processing techniques and mathematical methods,noise can be effectively suppressed but is limited by various factors.With the continuous development of deep learning techniques,generative adversarial networks are a powerful generative model that can learn speech data distribution and generate high-quality speech signals,making it a highly promising speech enhancement model.In this paper,we analyze and explore the speech enhancement method of generative adversarial networks.By analyzing the overall architecture of the model and the details of the network structure,we propose an improvement method and construct a multi-test set to verify the effectiveness of the algorithm by addressing the problems of insufficient feedback information of the existing model discriminator and insufficient ability of the self-attentive mechanism to handle long sequence numbers.The main work of this paper is as follows:(1)Analyze the key technologies of speech enhancement methods,conduct simulation experiments on subspace-based methods,spectrum mapping-based enhancement methods,and speech enhancement methods based on generative adversarial networks.Analyze the problems of each speech enhancement method to provide ideas for the following research.(2)To address the problems of insufficient discriminator feedback and insufficient analysis of human voice noise in the speech enhancement methods of generative adversarial networks,a dual discriminator generative adversarial network speech enhancement method is proposed.This is done by constructing two discriminators,constructing multi-type speech signal discriminant terms from the perspective of human voice noise speech emotion,designing the objective function for this purpose,and designing an adaptive optimizer for the characteristics of generative adversarial networks to improve the enhanced speech quality.At the same time,three test sets are designed based on the above method to verify the effectiveness of the method’s speech enhancement performance from multiple perspectives,with particular analysis on the performance in human voice noise environments.The experimental results show that the proposed method achieves high-quality speech enhancement.(3)To address the problems of low utilization of network parameters in the generative adversarial network speech enhancement method and the lack of consideration of location information in the self-attention mechanism,an attention mechanism for relative location is proposed.This is done by adding relative location information to the self-attention mechanism and applying the attention layer to the generative adversarial network model.The performance of the model can be improved by making better use of contextual information.Three test sets are designed based on the above method,and the effectiveness of the speech enhancement performance of the method is verified through multiple perspectives.The experimental results show that compared with the baseline method,the method achieves improvement in different signal-to-noise ratios and noise environments,realizes high-quality speech enhancement,and has strong speech enhancement and generalization capabilities.