Binaural speech separation based on the time-frequency binary mask

The perceptual ability of the human auditory system in capturing the target voice and filtering out the interferers has been remained as a great challenge. This paper proposes a binaural system for speech segregation based on spatial localization cues: Interaural Time Differences (ITD) and Interaural Intensity Differences (IID). A target speech signal is separated from interfering sounds by estimating time-frequency masks using the multi-level extension of the Otsu thresholding algorithm used in image segmentation. The ITD and IID are important features for mask estimation in low and high frequencies, respectively. A systematic evaluation in terms of Perceptual Evaluation of Speech Quality (PESQ) index shows that the resulting system yields significant improvement in performance of speech separation.