In audio signal processing, sound source localization (SSL) is a mature field. This project, however, looks at the SSL problem in a slightly different perspective, in which locations are computed softly. Specifically, instead of making point-decisions about the direction of arrival (DOA) of sound sources, a belief distribution about sources' locations over frequency and DOA/space is computed. The proposed approach generalizes the traditional SSL problem, and its outcome can also be used for other audio applications, such as spatial filtering. This project also investigates the SSL performance over a large set of microphone array configurations: a linear 4-element array (Kinect), a circular array with omnidirectional and cardioid microphones, and a 2-element, front-back array, with the goal of finding a "silver bullet" algorithm for all configurations. In addition to performance, computation cost is also accounted in the evaluation of different algorithms. Finally, evaluation is performed on a real data set collected at different SNR in a conference room.
