A FRAMEWORK FOR PRACTICAL ACOUSTIC SOURCE LOCALIZATION

An important aspect of audio-video conferencing and hands free telephony is the quality of the voice transmission. A microphone array (beamformer) can be used to improve voice quality, if the location of the talker is known. Determining the location of a talker, referred to as passive source localization, is a difficult problem and is the main focus of this thesis. The first step in passive source localization involves estimating the difference in signal propagation time to the various microphones. This is referred to as time difference of arrival (TDOA) estimation. The TDOA estimates are then used to estimate the source location, which involves solving a set of nonlinear equations. The thesis shows comparisons of generalized cross-correlation methods which are used to estimate the TDOAs. The method selected for estimating the TDOAs involves using the phase transform technique with weighted overlapped segment averaging. The TDOA algorithm also includes conditions on the signal power and the physical consistency of the estimates. The location of the audio sources were determined using iterative techniques that solved the nonlinear equations. Two techniques were compared - Powell's method and the simplex algorithm. The simplex algorithm was selected since it converged to the minimum of the error surface with far fewer iterations than Powell's method. Localization performance was evaluated using Monte Carlo simulations and real acoustic data which were obtained from two different four-element microphone arrays (rectangular arrays with 25 cm and 50 em microphone spacings). Generally, the localization of a specific source with the 50 em array yielded a lower localization error for a given TDOA estimate error than the 25 cm array. Localization error was also generally smaller for sources closer to a particular array than sources further away. Most of the localization error for a particular source away from an array manifested itself as error in the range from the centre of the array. Thus, the range was difficult accurately estimate, while the bearing (angle) could be accurately estimated