Attention driven computational model of the auditory midbrain for sound localization in reverberant environments

Abstract

In this paper, an auditory attention driven computational model of the auditory midbrain is proposed based on a spiking neural network [17] in order to localize attended sound sources in reverberant environments. Both bottom-up attention driven by sensors and top-down attention driven by the cortex are modeled at the level of an auditory midbrain nucleus - the inferior colliculus (IC). Improvements of the model in [17] is made to increase biological plausibility. First, inter-neuron inhibitions are modeled among the IC neurons which have the same characteristic frequency but different spatial response. This is designed to mimic the precedence effect [15] to produce localization results in reverberate environments. Secondly, descending projections from the auditory cortex (AC) to the IC are model to simulate the top-down attention so that focused sound sources can be better sensed in noise or multiple sound source situations. Our model is implemented on a mobile robot with a manikin head equipped with binaural microphones and tested in a real environment. The results shows that our attention driven model can give more accurate localization results than prior models.