Research On Algorithms Of Microphone Array Speech Enhancement

Speech is the most commonly used communication method.However,noise is everywhere around us.In most cases,noise will reduce the quality of speech,the efficiency of humancomputer interaction,and irritate hearing,annoy people.Therefore,the speech enhancement technology that can suppress the noise becomes very important.With the widespread use of microphone arrays,speech enhancement algorithms based on microphone arrays are receiving increasing attention.The research of this thesis is divided into two parts: minimum variance distortionless response(MVDR)and Wiener post-filtering in beamforming algorithms,and variable span linear filtering(VSLF)algorithm.The traditional MVDR speech enhancement algorithm uses the least mean square adaptive algorithm to update the weight vector,which can not track the changes of the statistical features of speech well,resulting in the general performance of speech enhancement.For this reason,this thesis draws on the idea of sampling matrix inversion(SMI)in narrow-band beamforming that updates the weight vector indirectly by updating the correlation matrix,but the updating rule is different from the SMI.The improved algorithm uses the different weight vector expression by transforming the noise correlation matrix into pseudo-coherence matrix.Diagonal loading techniques are also used to improve the robustness of the proposed algorithm.Experiments show that the improved MVDR algorithm outperforms traditional algorithms in both SNR and PESQ.Beamforming has limited ability to suppress non-coherent noise.It is often combined with post-filtering algorithm to further eliminate residual noise.The traditional Wiener postfiltering algorithm is based on the assumption of non-coherent noise field.In the actual environment,the coherence of noise in low frequencies is high,so the traditional algorithm has more noise residuals in low frequencies.In this thesis,the generalized prior power spectral density(PSD)estimation methods of desired signal and noise that not limited to a certain noise field,are used to construct a parametric Wiener filter,in which the parameter related with frequency and SNR can be adaptively adjusted according to different frequency bands and SNR to compensate for underestimated noise PSD,and then to improve noise suppression.Experiments show that the improved post-filtering algorithm can effectively suppress low-frequency residual noise,with improving SNR and PESQ.The beamforming algorithm is sensitive to DOA errors,so other multi-channel algorithms have attracted the attention of scholars.VSLF is such a kind of linear filtering algorithm combining the idea of subspace method.The correlation matrices of desired signal and noise are jointly diagonalized to obtain eigenvalues and eigenvectors.It is flexible to construct filters based on different optimal criterias by using different numbers of eigenvalues and eigenvectors,such as maximum SNR filter,MVDR filter and Wiener filter.In the second part of research,the VSLF algorithm is studied.The time-recursive averaging(TRA)method to estimate noise correlation matrix is proposed and the VS-MVDR algorithm belongs to VSLF is implemented.The algorithm itself has multiple variable factors,such as the number of microphones,filter length,and forgetting factor size.This thesis explores their effects on the performance of the algorithm under two different ideal conditions to guide the selection of parameters in the algorithm implementation process.Experiments show that the proposed TRA method performs better than the VAD-based noise correlation matrix estimation method.The implemented VS-MVDR can significantly suppress noise and improve speech quality.Finally,the algorithm performance is tested by actual recorded speech,and compared with the one in first part.The advantages and disadvantages of the two types of algorithms are briefly analyzed.Compared to the beamforming algorithm,VS-MVDR does not suppress enough coherent noise without utilizing the spatial information directly,but performs well in terms of distortion,and does not require further processing to eliminate non-coherent noise.