Research On Single Channel Speech Signal Separation Based On Sparse Representation And Deep Learning

Blind source separation refers to the process of recovering source signal from observed signal,called mixed signal as well,based on prior information of source signal when both source signal and mixing process are unknown or partial unknown.When the number of observed signal is one,it is called single channel blind source separation.At present,some achievements have been made on single channel blind source separation algorithm based on sparse representation,but there still exists some shortcomings.To solve this problem based on sparse representation involves the construction of redundant dictionary and the application of optimization algorithm,being of great theoretical significance.Meanwhile,single channel blind source separation is widely used in real life,so it is of great practical value to do some reasarch on this issue.In this paper,the theory of speech signal sparse representation and some existing algorithms to solve single channel speech signal separation based on sparse representation are explored in this paper.On this basis,the reasons for the problem of "cross projection" are analyzed.A method of constructing joint dictionary with common sub-dictionary is proposed,and a hybrid speech separation algorithm based on the joint dictionary is given.In this paper,we try to introduce the deep learning technology to this problem.Morever,two output hybrid speech separation model of heterosexual speaker based on deep neural network is established,good separation effect being obtained.The main research and innovative achievements of this paper can be seen as follows:(1)The sparse representation theory of signal is introduced and the basic principle of single channel blind source separation algorithm based on sparse representation theory is expounded.Moreover,the steps of speech signal preprocessing are detailed describled,and the evaluation criteria for measuring the quality of speech signal separation are introduced.On this basis,the experiment simulation is carried out to realizes the mixed speech separation of the heterosexual speaker based on sparse representation.(2)A method of solving single channel blind source separation based on joint dictionary including a common sub-dictionary is come up with.In this paper,the existing problems in the traditional way to solve single channel blind source separation based on sparse representation are explored.It is verified theoretically and experimentally that due to the similar components between dictionary training sets,resulting in the poor distinguishability of the identity sub-dictionary,which affects the separation effect.Based on this conclusion,a method to construct joint dictionary with a common sub-dictionary containing the similar components of identity dictionary is proposed.By introducing the common sub-dictionary,the similar components in the source signal can be projected on this common sub-dictionary,and the interference between those dictionaries is reduced at the same time,the problem of "cross projection" being overcome.And then,a detailed algorithm for solving blind source separation problem based on a joint dictionary with a common sub-dictionary is presented in this paper.Finally,the effectiveness of the new method is verified by experiments,and the influencing factors of the experimental results are analyzed.A detailed algorithm for solving blind source separation problem based on a joint dictionary with a common sub-dictionary is presented at the end of this part.Finally,the effectiveness of the new method is verified by experiments,and the influencing factors of the experimental results are analyzed.(3)A mixed signal separation model of heterosexual speaker based on deep learning is proposed in this part.Considering the real-time performance and quality of separated signal,the deep learning technology that is very popular and extensively used in the field of speech signal processing is applied.The theoretical knowledge of the deep neural network is introduced,including the structure of the deep neural network and the training algorithm of the network.A mixed speech separation model of heterosexual speakers based on depth learning is built in this section.What's more,the validity of the model is verified by experiments,and the influencing factors are analyzed.