Pitch recognition with sparse-coding recurrent neural networks

Abstract

Pitch is a percept of the mind that correlates to, but not solely with, acoustical periodicities. Simply stated, it relates different acoustic signals that share the same repetition rate into one percept. In contrast, various stimuli entering the acoustical periphery are expressed differently in the auditory nerves population response. Thus, there is a difference between the input activity and the percept of the brain. Broadly, two leading opponent approaches exist, temporal models and spatial models, but these models can explain partly disjoint features of the pitch. We propose a novel approach to bridge the gap between these two extremities. It is based on optimality constraints that assume parsimonious representation of the sensory auditory input. A recurrent neural network is trained to extract spatiotemporal patterns—pitch cues—from the auditory nerves population activities. These pitch cues are then linearly related to the stimulus' pitch. The model can explain different stimuli from psychoacoustic experiments, such as resolved and unresolved pitch, Transposed Tones, iterated rippled noise, nonlinear amplitude responses of the stimuli, harmonic shift, and musical notes. The uniqueness of this model is its ability to explain various psychoacoustic phenomena within one mathematical framework that is suggested to be shared with other modalities.