Representation of pitch input to neural network models of music perception

Abstract

Neural networks are a general class of computational models that can be used to model a variety of music perceptual tasks. An important issue in designing a neural network is the representation of input. The choice of representation can influence the network's trainability, its plausibility as a perceptual model, and its ability to generalize to other musical tasks. Networks that have been trained either to classify musical chords or to identify musical pitch are described. Four approaches to representation are examined. The simplest is a tone‐chroma notation in which there are 12 possible input nodes, one for each tone of the Western chromatic scale. Two approaches, a harmonic and subharmonic template, are motivated by theories of complex pitch perception. Input nodes are quantized into pitch‐class categories of the Western chromatic scale, and incorporate the notion of pitch height as well as tone chroma. In the fourth approach, each input node represents a frequency bin with a one‐third semitone bandwidth. This enables coding of input frequencies that are mistuned with respect to the standard tuning of Western music.