Spike-encoding of auditory stimuli : modeling studies of pitch and timbre

Abstract

Two studies are presented that relate psychoacoustic phenomena to neural coding mechanisms. The first describes a new method for estimating pitch from an autocorrelation-like representation of an auditory signal. Specifically, the autocorrelation representation is subjected to a smoothing operation before pitch estimation. This modification enables prediction of the pitch of a stimulus with a mistuned harmonic, a phenomenon that had previously presented a challenge to autocorrelation-based approaches. The resulting model retains the ability to predict many other pitch phenomena. A testable physiological mechanism is proposed that may underlie the pitch shift of a mistuned harmonic complex. This mechanism relies on the dynamics of synchrony-enhancing neurons in the cochlear nucleus, and it is demonstrated that a model of such a neuron, when responding to nearly harmonic stimuli, produces interspike intervals that dilate or contract in a manner that is qualitatively consistent with the pitch shift. The second study concerns the contribution of onset transients to the perception of timbre. It is proposed that information about stimulus onsets is encoded in the precise volley patterns of cortical onset responses. To test this idea, a model of the auditory periphery is fed into a reservoir of spiking neurons that has been designed to capture several important characteristics of a local cortical circuit: realistic spiking dynamics, conduction delays, a rich collection of recurrent connections, and synapses that adapt according to spike-timing-dependent plasticity. Correlational analysis of the model’s responses demonstrates that precise spike patterns corresponding to particular stimulus categories do develop, and spike-timing-dependent plasticity reinforces these patterns. A latency classifier that operates on this information is constructed, and its performance is evaluated against a related study. The spike pattern approach is seen to be more informative in many cases than previously published results. The generalization ability of the classifier is tested, and the results indicate that it is capable of generalizing well on new stimuli provided that onset characteristics are not disturbed. Finally, it is shown that the classifier relies on precise spike patterns rather than overall firing rates.