A pattern classification approach to sound source localization

One of the main limitations of traditional approaches to sound source localization is that they rely only on those wavefronts that are not subject to reflections. As a consequence, they may fail in case of unavailability of direct paths between the source and the microphones. In this paper we propose a pattern classification approach that takes advantage also of reflected waves by comparing Global Coherence Field or Oriented Global Coherence Field maps with a set of previously calculated models. The algorithm is able to estimate the source position as well as orientation, the latter in case a directional source is employed. Preliminary experiments conducted on both simulated and real data collections show that the algorithm can tackle cases in which reflections are predominant at each microphone. Moreover the proposed approach proves to be robust to slight deviations in the source position and orientation between the training and test data sets.