An Approach for Single-Channel Sound Source Localization

Sound source localization for machines has been studied in microphone array and binaural paradigms in most cases, while much less work has been done in the single-microphone or monaural paradigm. This paper addresses this task and presents a system designed to classify azimuths of a speech-emitting source with respect to a binaural receiver, however using only one of its ears. The system uses the spectrum second derivative approximation calculated on short duration frames and based on a bank of gammatone filters, in conjunction with a classifier artificial neural network. It is tested to explore its abilities and the influence of different parameters on its performances. True recognition rates and confusion matrices are reported in different evaluations studying the effects of the frame duration, filterbank size, silence elimination, generalization capabilities and source movement. Reported results show an ability to classify azimuths correctly up to a certain extent depending on the parameters used, with confusions occurring mostly with neighboring azimuths. The presented system can be built upon for more efficient localization of speech sources in both azimuth and elevation components.