Computational modelling of neural mechanisms underlying natural speech perception

Abstract

Humans are highly skilled at the analysis of complex auditory scenes. In particular, the human auditory system is characterized by incredible robustness to noise and can nearly effortlessly isolate the voice of a specific talker from even the busiest of mixtures. However, neural mechanisms underlying these remarkable properties remain poorly understood. This is mainly due to the inherent complexity of speech signals and multi-stage, intricate processing performed in the human auditory system. Understanding these neural mechanisms underlying speech perception is of interest for clinical practice, brain-computer interfacing and automatic speech processing systems. In this thesis, we developed computational models characterizing neural speech processing across different stages of the human auditory pathways. In particular, we studied the active role of slow cortical oscillations in speech-in-noise comprehension through a spiking neural network model for encoding spoken sentences. The neural dynamics of the model during noisy speech encoding reflected speech comprehension of young, normal-hearing adults. The proposed theoretical model was validated by predicting the effects of non-invasive brain stimulation on speech comprehension in an experimental study involving a cohort of volunteers. Moreover, we developed a modelling framework for detecting the early, high-frequency neural response to the uninterrupted speech in non-invasive neural recordings. We applied the method to investigate top-down modulation of this response by the listener’s selective attention and linguistic properties of different words from a spoken narrative. We found that in both cases, the detected responses of predominantly subcortical origin were significantly modulated, which supports the functional role of feedback, between higher- and lower levels stages of the auditory pathways, in speech perception. The proposed computational models shed light on some of the poorly understood neural mechanisms underlying speech perception. The developed methods can be readily employed in future studies involving a range of experimental paradigms beyond these considered in this thesis.

Publication
Thesis