Research And Implementation Of Adaptive Beamforming Algorithm Based On Blind Estimation And Eigen Decomposition

The multi-channel speech enhancement technology based on microphone array estimates the target and non-target components based on the multi-channel received signal's multiple characteristics such as time domain,frequency domain and space domain,and achieves the enhancement of target signal and the suppression of interference and noise through beamforming algorithm.Minimum Variance Distortionless Response beamforming is a common theoretical optimal beamformer,but in the process of practical application because of inevitable array model errors,resulting in serious degradation in performance.Therefore,based on the intelligent human-computer interaction noisy scene,this paper makes an in-depth study on the actual performance deterioration of the far-field small array caused by the measurement error of the array element and the error of the theoretical propagation model.Specific research contents are as follows:(1)An error correction algorithm based on blind estimation and eigen decomposition is proposed to solve the problem that the performance of MVDR beamforming algorithm is degraded due to array model error in practical application.The observation signals of microphone array are classified by cluster analysis,and the covariance matrix of target class and noise class is estimated.At the same time,the covariance matrix is trained in real time by adding the historical information of covariance matrix.The principal eigenvector is used as the unbiased estimation of the target signal steering vector.The cocktail party experiment shows that the error correction algorithm proposed in this paper can effectively avoid the influence of array model error and improve the speech enhancement performance and robustness of the beamformer.(2)Based on the proposed error correction algorithm and the MVDR beamformer architecture,an improved linear constrained adaptive beamforming algorithm is proposed to further improve the beamformer's interference and background noise suppression.The experiments show that the output SNR and SIR can be further improved by applying the zero constraint to the interference source and the white noise gain constraint to the background noise by combining the correction results of covariance matrix and steering vector.(3)An adaptive multi-channel real-time speech enhancement system based on ARM processor was built according to the engineering application and productization requirements of practical Intelligent speakers.The optimization strategy for real-time and robust performance of the system is designed and implemented,including: waken word detection based on speech signal energy and mark strategy based on signal energy;Acceleration strategy based on Floating Point Unit;An accelerating method based on advanced single instruction multiple data engine;Parallel programming technology based on Open MP.In addition,a two-stage adaptive sleep optimization strategy is designed to reduce the power consumption of the system.The experimental results show that the realtime optimization method can significantly reduce the system processing delay,and the low-power design can significantly reduce the system power consumption.In conclusion,this paper proposes an adaptive error correction algorithm based on blind estimation and eigen decomposition for the degradation of the practical performance of traditional beamforming algorithms.An improved linear multi-constraint adaptive beamforming algorithm is constructed and implemented for typical intelligent human-computer interaction noisy scenes such as far-field small array,which successfully improves the target enhancement effect,real-time performance and robustness of the beamformer in practical application.The research results of this paper provide some theoretical basis and guidance for the improvement of speech enhancement performance and system robustness in multi-channel speech enhancement applications such as intelligent human-computer interaction.