Research Of Adaptive Beamforming Algorithm Based On Blind Source Separation

Adaptive beamforming algorithm uses the time,space and frequency diversity of multichannel noisy signals received by microphone array to extract high-quality target speech signal and suppress interferences and background noise.The minimum variance distortionless response(MVDR)adaptive beamformer has been widely used for its optimal theoretical target enhancement performance.However,in the real multi-source dynamic acoustic scenes,the MVDR beamformer suffers from serious performance degradation due to the inevitable prior information errors,such as microphone array position measurement errors,signal propagation model mismatching,and source localization errors.Therefore,in order to overcome the degradation brought by the prior information errors,this thesis proposes an adaptive beamformer based on blind source separation algorithm(BSS)to significantly improve the beamforming performance in real application scenes.In addition,for the adverse situations with multi-source aliasing and dynamic scene mutation,adaptive error correction algorithms based on clustering analysis are proposed to improve the robustness and processing time of beamformer,providing feasible solutions for the small intelligent human-computer interaction devices.The contents are listed as follows:(1)An adaptive beamforming algorithm based on blind source separation is proposed to solve the problem of serious beamforming performance degradation in real cases,which is resulting from unpredictable prior information errors.This part includes blind source separation algorithm based on complex Gaussian mixture model,covariance matrix correction algorithm of non-target cluster based on speech existence probability,and target steering vector correction algorithm based on covariance matrix estimation.Experimental results show that proposed algorithms significantly improve the target Signal-to-Interference Ratio(SIR)in the real application scenes,even with unknown microphone positions,source type and acoustic environment.(2)For severe situations with dynamic sources,multi-source aliasing and scene mutation,adaptive error correction algorithms based on clustering analysis are proposed to improve the robustness and processing time of beamformer,including target cluster initial value optimization by time compression of keyword-interval-frame-discarding algorithm,the interference cluster initial value optimization by region scanning of signalframe-marking algorithm,the real-time sample complement method based on historical data fusion,and the iterative parameter fusion strategy based on time attention are proposed to deal with sudden interference.Experimental results show that the SIR of recovered target signals,real-time performance and robustness of beamformer are improved effectively.(3)In order to further suppress background noise and solve the problem of signal component whitening caused by blind source clustering,an adaptive time-frequency masking algorithm is proposed as post-processing model of beamformer.Experimental results show that the white noise suppressing performance of beamformer is improved effectively.To sum up,in order to solve the problem of speech enhancement performance deterioration caused by inevitable prior information errors in multi-source dynamic noisy scene,this thesis proposes an adaptive beamforming algorithm based on blind source clustering method,which can adaptively correct the estimation errors of target steering vector and covariance matrix,further significantly improve the practical performance of beamformer.In addition,for the multi-source aliasing and dynamic sever scenes,adaptive error correction algorithms based on clustering analysis are proposed to improve the robustness and processing time of beamformer.At last,an adaptive time-frequency masking algorithm is proposed as post-processing model to improve the background noise suppressing ability and solve signal whitening problem.The research results of this thesis provide strong theoretical and technical supports for the realization of small realtime speech enhancement devices.