Soft masking based adaptation for time-frequency beamformers under reverberant and background noise environments

We propose a new method to adapt a beamformer under reverberant and background noise environments. Several adaptation approaches have two stages consisting of time-frequency masking and an update of the beamformer. However, in these approaches, the adaptation error is large under such environments because the sparseness assumption does not hold. We focus on the premise that the adaptation error can be reduced by avoiding the degradation caused by the overlap between sources in time-frequency bins. Therefore, we derive a formula to update the beamformer in cases when overlap exists by using soft masking. The proposed method controls adaptation with soft masking to avoid the degradation caused by the overlap and reduces the adaptation error. Experimental results under a reverberant and background noise environment indicate that the proposed method improves the performance.