Co-channel Speech Separation Based on Amplitude Modulation Spectrum Analysis

A lot of effort has been made to achieve co-channel (two-talker) speech separation. However, the comprehensive analysis of the amplitude modulation spectrum (AMS) to address this problem has received little attention. In this paper, we propose an approach to exploit the AMS and to perform the separation based on the framework of computational auditory scene analysis (CASA). Specifically, this method utilizes the periodicity encoded in the AMS and then makes the channel selection. The main features of the approach are: (1) the reassignment method is used to improve the spectral resolution of the AMS in short duration; (2) a template-based pitch detector is used to determine the dominant fundamental frequency (F0) in an individual channel; (3) segmentation and grouping, the two stages in the CASA-based approaches, are employed to increase the robustness of channel selection. Systematic evaluation and comparison show that the proposed approach yields better performance than the previous system.