Research On The Improvement Of Speech Enhancement Algorithm Based On Sparse Representation And Dictionary Learning

In speech transmission,the normal reception of information was often interfered by noise.Therefore,speech enhancement plays an increasingly important role in signal processing.Because some noise is non-stationary,the speech enhancement effect of some traditional methods are not obvious.Using sparse representation and dictionary learning to process speech signals is undoubtedly an effective denoising method.Therefore,several improved speech enhancement methods based on sparse representation and dictionary learning are proposed in this paper.Dictionary learning and non-negative matrix factorization algorithm are two common speech enhancement algorithms,both of which are based on sparse representation and dictionary learning.In order to improve the efficiency of dictionary learning and solve the problem of large amount of sample training and high redundancy in sample,a fast dictionary learning algorithm based on the low-dimensional representation of signal subspace is proposed.The results show that,compared with KSVD algorithm,the fast dictionary learning algorithm achieves high noise reduction quality with faster dictionary training speed.The new approach has important research effect for signal processing.In order to reduce the coherence of the joint dictionary,an improved joint dictionary speech enhancement based on non-negative matrix factorization is proposed by enhancing the features of prior information structure.The results show that,compared with fast dictionary learning and K-SVD,the proposed algorithm with high speed of speech enhancement has a better SNR,which is better than the other two methods.Fast dictionary learning based on low dimensional representation of signal subspace can be divided into three stages: low dimensional representation of subspace,dictionary learning stage,and dictionary transformation stage.(1)Low dimensional subspace representation of data to be processed.The subspace low dimensional representation stage consists of principal component matrix acquisition and optimal feature extraction threshold selection.Firstly,the principal component analysis is carried out on the original data signal,and the optimal feature extraction threshold is selected by analyzing multiple indicators synthetically to construct the subspace.Then the low dimensional representation of the original signal is realized.(2)Transformation of subspace dictionary.The subspace dictionary is trained and the K-SVD dictionary learning algorithm is selected to generate the subspace dictionary.The orthogonal transformation matrix is constructed by analyzing the mathematical relationship between the subspace dictionary and the original dictionary.Then the subspace dictionary is carried out dictionary transformation to obtain original signal learning dictionary.The improved joint dictionary speech enhancement based on non-negative matrix factorization is mainly divided into joint dictionary training stage and speech enhancement stage.(1)Composite processing.In the stage of constructing joint dictionary,the training data is composite processed without introducing any priori model with high computational complexity.The feature of prior information structure is enhanced by selecting appropriate expansion step,then the structure of speech and noise signal is more obvious,the discriminability of the basis vector is improved to constrict the low coherence joint dictionary.(2)Speech enhancement.In speech enhancement stage,the signal with noise is composite processed by selecting expansion step and is sparse represented on the low coherence joint dictionary.The pure speech signal is obtained by joint dictionary and coefficient matrix extracted by information separation.The results show,the novel method achieves better speech enhancement effect.