Soft-CASA system for single channel speech separation

Abstract

In this paper we study the masking effect on Computational Auditory Scene Analysis (CASA) based systems for single channel speech separation (SCSS). In this study, we focus on the benchmark masks of the literature that are namely: the ideal binary mask (IBM), the binary mask (BM) and soft mask. Each system is evaluated objectively and subjectively in order to highlight the effect of each mask on the intelligibility and the quality of the separated speech. Based on this study we develop a new system, that we call Soft-CASA for SCSS that outperforms the original one. The proposed system achieves 28.84% improvement in the Short-time Objective Intelligibility (STOI) parameter, 7.1% improvement in SNRloss and 92% improvement in overall perceptual score (OPS), compared to the original system.