Low Rank Sparse Decomposition Speech Enhancement Method Based On Constraints Of Continuity

As an important branch of speech signal processing,speech enhancement technology has a wide range of applications in many areas of real life.However,speech signals are often affected by interference signals.Under normal circumstances,these interference signals have random characteristics.Therefore,it is difficult to completely filter out noise and obtain pure speech.Therefore,finding an efficient speech enhancement method for noise suppression processing of noisy speech signals has important research value for improving the enhanced speech quality and intelligibility and improving the anti-noise effect of the speech system.The constrained low-rank sparse matrix factorization(CLSMD)method ignores the temporal continuity property between adjacent speech frames during speech enhancement,resulting in decomposition of sparse matrices with distinct discrete outliers.Therefore,in order to improve the noise suppression capability and improve speech quality and intelligibility of the speech system,this article proposes an improved speech enhancement method,temporal continuity constrained low-rank sparse matrix decomposition(TCCLSMD)method and a temporal continuity constrained nonnegative low rank sparse decomposition(TCNLSMD)method.In both methods,in addition to adding low-rank sparse constraints,temporal continuity characteristics constraints are added.After the sparse matrix is generated by the singular value decomposition(or eigenvalue decomposition)and the hard threshold function of the test matrix in this paper,guided sparse component reconstruction reduces(or eliminates)the isolated discrete points under the constraints of time continuity,and finally achieves speech enhanced purpose.In a variety of realistic noise environments,the two methods proposed in this paper were tested experimentally and compared with the current mainstream speech enhancement methods.The experimental results show that under various types of noise testing conditions,these two methods improve the noise suppression capability compared to the current mainstream speech enhancement methods,leaving less residual noise and improving the quality of enhanced speech.The methods and work innovations proposed in this topic are as follows:1、A speech enhancement method based on TCCLSMD.On the basis of the sparse matrix obtained by singular value decomposition and hard threshold function estimation,the sparse matrix is reconstructed by adding temporal continuation constraints to reduce discrete outliers,preserve more speech information,and reduce enhanced speech distortion.2、A speech enhancement method based on TCNLSMD.On the basis of sparse matrix obtained by eigenvalue decomposition of matrix and hard threshold function estimation,reconstruction of sparse matrix is guided by adding temporal continuation constraints to improve system noise suppression capability and improve enhanced speech quality and intelligibility.