Subband beamformer combined with time-frequency ICA for extraction of target source under reverberant environments

We propose a novel beamforming method that is complementarily combined with the time-frequency domain independent component analysis (TF-ICA). Under reverberant environments, performance of TF-ICA is known to be limited. On the other hand, when time period in which there are only noise sources (noise-only period) is given, beamforming based on multichannel spatial prediction (MSP-BF) are shown to achieve good noise reduction performance. The proposed method applies noise-period identification based on the output of TF-ICA to obtain a noise-only period. MSP-BF using the obtained noise-only period is performed. Furthermore, to avoid performance degradation, the proposed method selects the better output signal obtained by TF-ICA and MSP-BF. Experimental results under a reverberant environment indicate that the proposed method outperforms TF-ICA. Furthermore, thanks to output-signal selector, even when there are no noise-only periods, performance degradation can be avoided.