ISCA Tutorial and Research Workshop on Statistical and Perceptual Audition (SAPA2006)
Pittsburgh, PA, USA
Most algorithms based on Computational Auditory Scene Analysis (CASA) for binaural speech separation do not have the ability to inhibit already localized and for a long time present sources in the auditory scene. This has the major drawback that the auditory cues of weaker and new sources are subject to interference from already localized and perceived signals and the separation performance is worse if the signals overlap in their processing domain. In this paper we outline how one can build intuitively a separation system that has this inhibition feature. The main block and starting point of our derivation is a simple cross correlation based localization system with two microphones. The inhibition is achieved by feeding back localization results to a filter and sum structure that cancels localized sounds. Interestingly, our intuitive approach leads to a special case of a well known time domain blind source separation algorithm which was derived from a statistical signal processing viewpoint and exhibits good convergence even in reverberant environments. Finally, we discuss how the insights gained from building a blind source separation this way can be used to integrate CASA techniques.
Bibliographic reference. Schölling, Björn / Heckmann, Martin / Joublin, Frank / Goerick, Christian (2006): "Structuring time domain blind source separation algorithms for CASA integration", In SAPA-2006, 37-41.