Generalization of supervised learning for binary mask estimation

This paper addresses the problem of speech segregation by estimating the ideal binary mask (IBM) from noisy speech. Two methods will be compared, one supervised learning approach that incorporates a priori knowledge about the feature distribution observed during training. The second method solely relies on a frame-based speech presence probability (SPP) es-timation, and therefore, does not depend on the acoustic condition seen during training. We investigate the influence of mismatches between the acoustic conditions used for training and testing on the IBM estimation performance and discuss the advantages of both approaches.