Beamforming array technique with clustered multichannel noise covariance matrix for mechanical noise reduction

Masahito Togami, Takashi Sumiyoshi, Yasunari Obuchi, Yohei Kawaguchi, Hiroaki Kokubo

Abstract: In this paper, we propose a novel multichannel noise reduction method for a mechanical noise with a time-variant impulse response. The mechanical noise source location moves depending on the status of the actuator. In accordance with the move of the noise-source location, a most suitable multichannel beamformer is selected separately at each time-frequency bin. Each multichannel beamformer is made from a corresponding multichannel noise co-variance matrix which is learned in advance. The selection criteria in the proposed method is to minimize the residual noise power in the output signal. The multichannel beamformer that minimizes the residual power after beamforming is selected. Furthermore, to reduce directional noise sources, the multichannel directional noise covariance matrix is inserted into each multichannel mechanical noise covariance matrix. Experimental results of mechanical noise reduction show that the proposed method can reduce the mechanical noise more accurately than the conventional method.