Multi-Agent Based Harmonic Stream Segregation for Auditory Scene Analysis

Abstract

We propose a new approach to the sound stream segregation based on the multi-agent paradigm. The sound stream segregation is a technology to extract individual sounds from a mixed sound. And it is considered as a building block for Auditory Scene Analysis, which provides a mechanism of understanding the acoustic environment. But it is difficult to completely analyze all the sounds in the real world. Instead, we propose the emergent computational model as a framework to actively process focused sounds. In this paper, we discuss stream segregation agents that focus harmonics. First, we explain a basic framework, in which each tracer agent extracts a sound using local information, and these agents interact to resolve sound overlaps. Next, we construct an advanced system by introducing new agents that control the actions of tracer agents. Experimental results show that the proposed system segregates sound streams stably and adaptively to the change of the input sound.