A delay-cell network for pitch estimation

Abstract

Autocorrelation functions have played a prominent role in many theories of pitch perception. A problem with relating such theories to a physiological substrate is that the autocorrelation function must be computed for delays as large as 5–20 ms which is much longer than what would be expected from conduction or synaptic delays. The delay-cell network [D. C. Mountain, ARO Abstracts 16, 107 (1993)] implements these delays via the combination of a fast-acting inhibitory input and a slower-acting excitatory input synapsing on the same cell (delay cell). The output from a layer of these delay cells (each with different delay) is then fed to an output layer consisting of coincidence detectors. This layer compares the delayed versions of the signal to the undelayed signal. The representation produced by the network is a smoothed interval histogram rather than an autocorrelation function. Simulations of the delay-cell network predict that it should be possible to elicit pitch percepts using click trains with a random interclick interval. Preliminary experiments indicate that these random click trains do elicit a pitch sensation which varies with the mean interclick interval. The autocorrelation functions for such stimuli are quite broad and exhibit no clear peaks. [Work supported by ONR.]