Study On Beamforming Based Speech Separation And Acoustic Echo Cancellation For Microphone Array

With the rise of various smart devices,speech recognition has become one of the important ways for human-computer interaction.In order to improve the accuracy of speech recognition in far-field situations,speech enhancement processing is usually required.At present,important advances have been made in speech enhancement techniques.A variety of speech denoising methods have been developed and widely used in practice.However,when the interference source is speech,these denoising methods will fail.Blind source separation technology can handle voice interference,but there are still some problems to be solved.This paper studies the sound source separation technology,firstly using the spatial characteristics of the microphone array(ie,the acoustic transfer function)to separate and enhance the speaker's speech signal,and using the speaker power spectral density matrix to construct the post-wiener filter to remove residual noise.This article studies the speech separation and echo cancellation techniques.The main contents are as follows:(1)The method of sound source separation using linearly constrained minimum variance(LCMV)Beamforming is studied,and its noise power spectral matrix estimation method is improved to suppress interference speech to a greater extent.On this basis,cascading post-wiener filtering is used for denoising,and the method of coefficient equalization is used to reduce the damage of target speech.(2)In the algorithm of sound source separation,it is necessary to use speech segments where each speaker speaks alone.For this reason,this article proposes a feasible method for capturing single source segment using the speaker segmentation for the situation in which two speakers speak simultaneously.(3)Improving the acoustic echo cancellation(AEC)module in WebRTC.Under double-talk conditions,the adaptive normalization and regularization coefficients are applied to improve the NLMS filter.Thus the effect of echo cancellation is improved.