Separation Of Overlapping Speech Based On Computational Auditory Scene Analysis

Auditory scene analysis is the study on how to recognize environmental events with sounds arising from various acoustic sources. When various acoustic sources exist simultaneously, the problem that one separates acoustic information derived from each acoustic source is called acoustic source separation problem. It is the fundamental technology to realize auditory scene analysis on computers.In this thesis, an effective system is described, which is successful in separating overlapping acoustic signals. It is known there are a lot of factors that influence segregation perception or fusion perception of frequency components. To perform acoustic source separation, it is necessary to investigate the principles under which human being extracts acoustic information belonging to respective perceptual sounds from a sequence of spectra.The main aspects of this thesis are given as following: (1) A new model simulating auditory system of human being is presented, by which the incoming array of sound are analyzed into a large number of frequency components.(2) The data coming from (1) is organized into discrete elemental units. Then we collect together the appropriate subsets of the elements that appear to come from single source. The output of this stage is one or more groups of the elements.(3) We can resynthesize or reconstruct the source signal using each group of element from stage(2).The experiment results indicate that the approach described here is efficient.