Adaptive sound localization with a silicon cochlea pair

A neuromorphic sound localization system is presented. It employs two microphones and a pair of silicon cochleae with address event interface for front-end processing. The system is based the extraction of interaural time difference from a far-field source. At each frequency channel, a soft-winner-takes-all network is used to preserve timing information before it is processed by a simple neural network to estimate auditory activity at all bearing positions. The estimates are then combined across channels to produce the final estimate. The proposed algorithm is adaptive and supports online learning, enabling the system to compensate for circuit mismatch and environmental changes. Its localization capability was tested with white noise and pure tone stimuli, with an average error of around 3° in the -45° to 45° range.