The Pattern-Transformation Model of Pitch

Abstract

This report describes a new approach to pitch perception, based on a kind of auditory pattern recognition. The general approach is formalized in a mathematical model, the “pattern-transformation model.” In the first stage of this model, an acoustic stimulus is assumed to be “transformed” by the sense organ into what is called a neural activity pattern. This pattern is thought to represent roughly the power spectrum of the waveform. Thus, the model is phase insensitive; the temporal fine structure of the stimulus waveform is ignored. The first-stage pattern (power spectrum) is then assumed to be Fourier transformed (neurally) into another pattern of neural activity. This second pattern roughly represents the autocorrelation function of the stimulus. Pitch is derived from the positions of maximal activity in this pattern. From preliminary tests, it appears that the model can accurately predict the pitches of many types of acoustic stimuli. In addition, the model provides numerical estimates of pitch “strength” and “ambiguity.” While experimental confirmation of these predictions is lacking, the strength estimates agree at least qualitatively with available data. [Research supported by a National Science Foundation Postdoctoral Fellowship awarded to the author and the Institute for Perception—TNO, Netherlands.]