Ideal binary masking in reverberation

The ideal binary mask has been established as a computational goal of binary time-frequency masking approaches to sound separation through experiments that show large speech intelligibility gains for both normal-hearing and hearing-impaired listeners. The ideal binary mask is commonly defined using a signal-to-noise ratio threshold for binary decisions called the local criterion. Recently the ideal binary mask definition was extended to deal with reverberant signals by introducing the reflection boundary, a division point between early and late reflections, which allows target reflections to be treated as either part of the desired signal or as noise. The current study refines the conclusion in previous work by analyzing the effects of both the reflection boundary and local criterion parameters. Experimental results show that ideal binary masks defined with reflection boundaries of 100 ms or less can produce significant intelligibility improvements, which establishes that binary masking can be effective for both noise reduction and dereverberation. Further, results show that to achieve intelligibility gains, early reflections should be preserved by the binary mask while late reflections should be treated as noise.