Physiology-based model of multi-source auditory processing

Abstract

Our auditory systems are evolved to process a myriad of acoustic environments. In complex listening scenarios, we can tune our attention to one sound source (e.g., a conversation partner), while monitoring the entire acoustic space for cues we might be interested in (e.g., our names being called, or the fire alarm going off). While normal hearing listeners handle complex listening scenarios remarkably well, hearing-impaired listeners experience difficulty even when wearing hearing-assist devices. This thesis presents both theoretical work in understanding the neural mechanisms behind this process, as well as the application of neural models to segregate mixed sources and potentially help the hearing impaired population. On the theoretical side, auditory spatial processing has been studied primarily up to the midbrain region, and studies have shown how individual neurons can localize sounds using spatial cues. Yet, how higher brain regions such as the cortex use this information to process multiple sounds in competition is not clear. This thesis demonstrates a physiology-based spiking neural network model, which provides a mechanism illustrating how the auditory cortex may organize up-stream spatial information when there are multiple competing sound sources in space. Based on this model, an engineering solution to help hearing-impaired listeners segregate mixed auditory inputs is proposed. Using the neural model to perform sound-segregation in the neural domain, the neural outputs (representing the source of interest) are reconstructed back to the acoustic domain using a novel stimulus reconstruction method.