A neural network model of musical chord classification

Abstract

A neural network, which models a musical task, specifically the chord classification task, was built. The model “listens” to a chord and classifies it as a major, minor, or diminished chord. Two basic approaches were taken in the development of the model: (1) pitch class, and (2) harmonic template. The pitch class approach represents chords as they are customarily represented in music notation, namely by specifying the pitch class of each tone in the chord. The harmonic template approach defines each pitch in terms of an equivalent harmonic complex, specified by the most appropriate pitch class label for each of the first five harmonics. This approach is motivated by pitch perception theories based upon pattern matching (Goldstein, 1973; Terhardt, 1974). Connection strengths between nodes in the network were derived using an error propagation learning algorithm (Rumelhart, Hinton, and Williams, 1986). The pitch class network performed poorly by classifying only 33%–72% of the musical chords correctly. The harmonic template network classified 100% of the musical chords correctly. [Work partially supported by the University of Washington Graduate School Research Fund and ONR Grant N00014-86-C-0065.]