A dynamical pitch model constructed from arrays of adaptive elements

Abstract

A dynamical model of pitch perception is presented. The model is constructed from parallel arrays of tuned, nonlinear, adaptive elements. An outer layer of filters and phase-locked loops provides continuous estimates of the dominant frequencies of the input. The outputs of the phase-locked loops are then processed by an array of adaptive pitch templates. The template system is a dynamical implementation of a classical approach to assigning pitch to harmonic (or quasi-harmonic) signals. It is shown the templates are capable of modeling a number of nonlinear pitch phenomena, and pitch-shift effects are given particular attention. The templates also provide a straightforward way to identify multiple simultaneous pitches and assign relative strengths to them. A neurocomputational justification for the action of these templates, in the form of a simple spiking neuron model, is presented. This model suggests that the harmonic grouping performed by the templates can be achieved through a temporal calculation on information derived from different locations on the tonotopic axis. After establishing the theoretical capabilities of these templates, the entire model is run on sample stimuli. The model is able to generate correct pitch identifications, but improving the tuning of the peripheral elements will be necessary before testing more complicated stimuli.