Dual-microphone based binary mask estimation for robust speaker verification

Missing feature theory (MFT) has shown great potential for robust speaker recognition in noisy environments. Accurate estimation of binary mask is crucial in MFT-based speaker recognition. This paper addresses the speaker verification problem using MFT in a practical scenario: the location of target speaker is fixed while the locations of noise interferences are unknown. Specifically, we propose a dual-microphone semi-blind approach to estimate the binary mask. During system initialization, a spatial location model for the target is trained precisely. Then a spatial model for corrupted speech is obtained on-line by model adaptation. Finally, the binary mask is estimated by likelihood comparison. Moreover, we propose a reliable frame selection method to further focus on the reliable speech frames for missing data speaker recognition. Experimental results demonstrate that our proposed approach achieves substantial improvements in recognition performance in both white noise and speech corrupted conditions.