Neural Basis and Computational Strategies for Auditory Processing

Abstract

Our senses are our window to the world, and hearing is the window through which we perceive the world of sound. While seemingly effortless, the process of hearing involves complex transformations by which the auditory system consolidates acoustic information from the environment into perceptual and cognitive experiences. Studies of auditory processing try to elucidate the mechanisms underlying the function of the auditory system, and infer computational strategies that are valuable both clinically and intellectually, hence contributing to our understanding of the function of the brain. In this thesis, we adopt both an experimental and computational approach in tackling various aspects of auditory processing. We first investigate the neural basis underlying the function of the auditory cortex, and explore the dynamics and computational mechanisms of cortical processing. Our findings offer physiological evidence for a role of primary cortical neurons in the integration of sound features at different time constants, and possibly in the formation of auditory objects. Based on physiological principles of sound processing, we explore computational implementations in tackling specific perceptual questions. We exploit our knowledge of the neural mechanisms of cortical auditory processing to formulate models addressing the problems of speech intelligibility and auditory scene analysis. The intelligibility model focuses on a computational approach for evaluating loss of intelligibility, inspired from mammalian physiology and human perception. It is based on a multi-resolution filter-bank implementation of cortical response patterns, which extends into a robust metric for assessing loss of intelligibility in communication channels and speech recordings. This same cortical representation is extended further to develop a computational scheme for auditory scene analysis. The model maps perceptual principles of auditory grouping and stream formation into a computational system that combines aspects of bottom-up, primitive sound processing with an internal representation of the world. It is based on a framework of unsupervised adaptive learning with Kalman estimation. The model is extremely valuable in exploring various aspects of sound organization in the brain, allowing us to gain interesting insight into the neural basis of auditory scene analysis, as well as practical implementations for sound separation in ``cocktail-party'' situations.