Researches On Signal Separation Based On Deep Learning Method

In the real environment,the speech signal that people are interested in is usually interfered by noise or other background sound,which seriously damages the quality of speech and reduces the performance of speech recognition.In order to distinguish the noise in real speech from the mixed target speaker's voice,speech separation technology is the most common method.Recently,due to the development of deep learning,speech separation technology has made great progress,but there are still many difficulties.Therefore,based on the above background,this thesis focuses on the research of independent component analysis,speech and video processing,generative adversial method and gradient constraint strategies.In the main part of the thesis,based on the deep learning method,the thesis mainly solves the problem of separating the independent components in the mixed speech signal to get the pure target speech.In view of the above problems,this thesis proposes the following new application schemes based on the method of independence analysis and correlation inhibition:(1)This thesis proposes a single channel speech separation structure based on independence analysis method and a single channel speech separation structure based on correlation inhibition method.In the structure of the independent analysis method,the structure obtains joint sampling and edge product sampling of mixed signals through separation and resampling modules.The idea of adversial network is used to optimize the similarity between the two samples to ensure that the separated signals are independent of each other;In the structure of correlation inhibition method,this structure starts from the perspective of minimizing the correlation between component signal layers,and fuses a new gradient constraint loss in the baseline model.It pays attention to the non-interference of the information distribution of edge gradient between signals,which makes the separation task get better effect in the optimization process.(2)This thesis proposes a speech separation structure based on frequency domain independent analysis method and a speech separation structure based on frequency domain correlation suppression method,combining image and speech as input.In the structure based on the independent analysis method of frequency domain,this structure makes use of the idea of adversial game,and use the separation and resampling modules to keep the independent relation between the separated signals.In the structure of the correlation suppression method based on the frequency domain,the method of constraining the correlation of each component signal in the representation space of the frequency domain ensures that the edge information between the component signals has enough differentiation.The essential structure is not changed in the process of separation,so as to improve the precision of the video and speech separation task with various background sounds.