Binary Mask Criteria Based on Distortion Constraints Induced by a Gain Function for Speech Enhancement

Large gains in speech intelligibility can be obtained using the SNR-based binary mask approach. This approach retains the time-frequency (T-F) units of the mixture signal, where the target signal is stronger than the interference noise (masker) (e.g., SNR > 0 dB), and removes the T-F units, where the interfering noise is dominant. This paper introduces two alternative binary masks based on the distortion constraints to improve the speech intelligibility. The distortion constraints are induced by a gain function for estimating the short-time spectral amplitude. One binary mask is designed to retain the speech underestimated (T-F) units while removing the speech overestimated (T-F)units. The other binary mask is designed to retain the noise overestimated (T-F) units while removing noise underestimated (T-F) units. Listening tests with oracle binary masks were conducted to assess the potential of the two binary masks in improving the intelligibility. The results suggested that the two binary masks based on distortion constraints can provide large gains in intelligibility when applied to noise-corrupted speech.