Speech intelligibility of ideal binary masked mixtures

An analysis of intelligibility measurements of ideal binary masked speech in noise for a group of normal hearing listeners is presented. In the proposed model, speech cues in the processed mixtures are encoded by two information channels: a noisy speech channel and a vocoded noise channel. Results indicate that the former dominates for dense binary mask patterns, and the latter for sparse binary mask patterns, as controlled by a local SNR criterion used for forming the ideal mask. Moreover, speech cues from the target part of the processed mixture may be better utilized by the listeners as a result of the ideal binary masking. Finally, the analysis is extended to show a good qualitative agreement with several previous studies of intelligibility of ideal binary masked noisy speech.