Modelling the effect of the missing fundamental with an attractor neural network

Abstract

In order to investigate the use of attractor neural networks (ANNS) in modelling real phenomena of human perception, we developed an ANN model of spectrum-invariant pitch recognition that was aimed at cortical mechanisms of pitch perception, and tested the model in computer experiments which were arranged to be compatible with three pitch-identification psychophysical experiments with human subjects. The results of computer and actual experiments are qualitatively the same, and quantitatively they depart at most by 10%, which can be considered to be in good agreement. It is shown that the present ANN model can provide an explanation for the different efficacies of lower and upper harmonics in conveying the missing fundamental pitch percept, by means of one universal cortical mechanism. In our model, a complex tone evokes the pattern of activity that consists of a particular combination of ‘isofrequency’ stripes in the ANN that represents the relevant part of the primary auditory cortex (AI). Recognition of the pitch is interpreted as an autoassociative retrieval of the corresponding stripe-like template. The numbers of active neurons in the stripes that represent lower frequencies are greater than the numbers of active neurons in the stripes for the higher frequencies. Thus, when the network is presented with the frequency complex containing either lower or higher harmonics, different efficacies of lower and upper harmonics in conveying the missing fundamental pitch percept are manifested. The proposed ANN model can be easily generalized for the spectrum-invariant recognition of unharmonic sounds, such as, for instance, speech.