In this thesis, an improved robust sound source localization (SSL) method based on the arrival time delay estimation is proposed for intelligent service robots. SSL systems estimate an azimuth of the sound source with two or more microphones. Conventional SSL methods are intensity based, time delay of arrival (TDOA)-based, and beamforming ones. Among conventional methods, intensity-based SSL methods usually fail to detect the accurate azimuth mainly because of the microphone gain calibration problem. An SSL method with differenced signals and the GCC-PHAT-based method are typical TDOA based ones. The difference-based SSL algorithm makes a contour by taking the difference between samples before estimating time delays. This one is an effective method when sound sources have rather unclear peak points. However, the performance degrades in noisy or reverberant conditions. The generalized cross correlation-based SSL algorithm estimates accurate azimuths by normalizing the frequency distortion using a weighting function in highly reverberant conditions. However, the error rate of the azimuth estimation is higher than other methods in mild reverberant conditions.
In this thesis, an SSL algorithm robust to microphone gains is proposed. The proposed SSL algorithm is robust to microphone gains because it uses only the time differences among microphones. To make it possible, a cost function which normalizes the microphone gains is utilized and a procedure to detect the rough position of the sound source is proposed. The performance of the azimuth estimation is improved 60% compared with that of the intensity-based method.
Other proposed methods such as adaptive mode selection, signal interpolation, and confidence measuring are suggested to improve the overall performance. Proposed adaptive mode selection unifies several SSL methods using the kurtosis measure. It can select SSL methods according to the types of signals. To improve the performance of the sound source ...