Enhanced audio source separation in room acoustic environments with selected binaural cues

We study the problem of audio source separation from stereo-channel microphones in a room acoustic environment. Instead of using the whole binaural cues as done in DUET (Degenerate Unmixing Estimation Technique)-type methods, we propose a technique that selects a reliable subset of binaural cues by examining the phase determinacy and the sparse source conditions. We conduct experiments with a simulated room acoustic environment and show a significant performance gain of the proposed technique over the DUET-type methods.