Localization of multiple sound sources in three dimensional environments

Abstract

Localization of sound sources has several applications like teleconferencing, speechrecognition, speaker identification, speech acquisition in an automobile environment,sound capture in reverberant enclosures, large room recording-conferencing, and hearing aid devices. One way of finding the location of a sound source is to utilize the direction ofarrival (DOA) values. This value indicates the angle between two lines, first of whichconnects the mid point of the microphone array and the sound source, and second of whichcarries the microphone array. DOA values are usually estimated by first estimating the time delay of arrival (TDOA) value of the signals received from two microphones and thenconverting TDOA estimates to DOA estimates. TDOA value is estimated using the crosspowerspectrum phase (CSP) coefficients.The existence of multiple sound sources in an environment brings two complications to this problem. First of all, the correlation between the sound sources distorts the TDOAestimation procedure. Second, multiple DOA estimates, which are calculated for multiplemicrophone arrays, have to be matched to the sound sources to find the correct intersection points among multiple ones. Another major complication is the generalization of the environment to the three dimensional space.In this thesis, a new method is proposed, for localizing multiple sound sources inthree dimensional environments. The synchronous addition of CSP coefficients method isutilized for finding the undistorted DOA estimates. Then these estimates are clusteredusing a new design-specific, an inconsistency measure based clustering algorithm. Having found DOA estimate triples for each sound source, the location of sound sources aredetermined by finding the intersection point of three cones formed by three DOA values.This intersection is found by first finding a closed formula, which consists of a single variable, for a locus of the intersection, which is a three dimensional path and then findingthe suitable point on this path. Experiments are done on simulation and real acoustical environments. The results are promising considering the complexities that the algorithm faced, and the number of themicrophones that are available for the experiments, and the diversity of them.