Research On Underdetermined Speech Blind Separation Based On Sparsity

Blind source separation technology is a powerful signal processing technology.It can separate and recover individual source signals only through several observation signals when the source signal and mixing parameters are unknown.Many mature theories and methods have been developed for(over-)determined separation problem.However,the problem of blind source separation under underdetermined conditions is more challenging and practical,causing the interest of many researchers and gradually developing into a new research hotspot.This paper proposes a new mixing parameter estimation method under the framework of the mainstream "two-stage" method for underdetermined blind source separation.At the same time,it improves the classical source-signal separation method based on time-frequency mask and apply it to underdetermined speech separation problem.Firstly,based on the principle of constructing smooth two-dimension histogram for mixing parameter estimation,this paper proposes a new mixing parameter estimation method combined with the watershed algorithm,which can solve the two major problems of source signal quantity enumeration and mixing parameter estimation at the same time.Compared with the traditional method of estimating the mixing parameters based on the clustering method,it has the advantages of fast operation speed and strong anti-interference ability.Secondly,the advantages and disadvantages of the two main source signal separation methods for time-frequency masking and sparse representation are analyzed and studied.Combining the advantages of both,an improved strategy is proposed based on the time-frequency masking method,which improves the algorithm to some extent.Finally,the effectiveness of the proposed algorithm is verified by designing related speech separation experiments.In this paper,the underdetermined speech separation problem is taken as the research object.Under the guidance of the mainstream "two-stage" method framework,the corresponding improvement strategies are proposed for the mixing parameter estimation step and the source signal separation step,respectively.Finally,this paper proposes a speech blind separation method with fast computation speed,good separation performance and strong robustness,and designs several related experiments to verify it.