Study On Adaptive Beamforming And Post-filtering Techniques For Microphone Array

Microphone array is widely used in hands-free communications and human-machine interfaces,and it can be a good solution to noise reduction,dereverberation,source localization or tracking,source separation and other issuses.As a result of the use of multiple microphone sensors,microphone array can not only use the time and frequency domain information,but also use the spatial information to process signal.The signal-to-interference ratio can be significantly improved by using the microphone array adaptive beamforming algorithm to steer the beam to the desired sound source or to nullify the interference direction.By utilizing the spatial selectivity,even the highly non-stationary interference can be effectively suppressed.However,the performance of the beamforming algorithm is limited for the uncorrelated noise,so the post-filtering algorithm shall be used to enhance the output signal of the spatial filter.In this thesis,the adaptive beamforming and post-filtering techniques based on microphone array are studied.The main works are summarized as follows.Firstly,it is difficult to accurately estimate the DOA in the environments where the signal-to-noise ratio is relatively low and the reverberation is relatively high.And,since the conventional post-filter design only considers the diffusion noise or only the single-channel output from the beamformer,we derived the steering vector estimation and the estimation of noise power spectral density based on eigenvalue decomposition.The estimation of the steering vector based on the maximum eigenvector does not need the geometric position information of the microphone array,and it is robust to the reverberation in environments.The estimation of the noise power spectrum density based on the minimum eigenvector is less affected by target signal leakage,which makes the noise estimation more accurate.The simulation results show that the proposed algorithm can better suppress the noise and reduce the distortion of the desired signal.Secondly,the generalized sidelobe canceller based on the recursive least squares algorithm with fixed forgetting factor is difficult to meet the requirement of fast tracking performance.In this thesis,we propose a variable forgetting factor based on the approximate derivative of filter coefficients to improve the tracking performance of adaptive algorithms.In order to further remove the adaptive beamforming residual noise,we deduce the cross-correlation spectral subtraction algorithm based on variable spectral subtraction factor.The variable spectral subtraction factor is based on the optimal priori SNR estimation.Compared with the traditional cross-correlation spectral subtraction method,the proposed algorithm is more capable of suppressing residual noise.The simulation results show that the proposed variable forgetting factor algorithm has faster convergence speed and better tracking ability,and the derivation of the variable spectral subtraction factor algorithm can effectively suppress the residual noise and increase the intelligibility of speech.