Neural Networks and Musical Style Recognition

3rd year dissertation project for a BSc in Physics and Music, University of Wales, Cardiff, 1998.

Abstract:

This project attempts to discover if a neural net is capable of performing stylistic discrimination between two pieces of music by two different composers, JS Bach and Igor Stravinsky. It was found that, although the trained network was able to discriminate between two different pieces, that a large number of samples from different pieces only served to confuse the network, resulting in it categorising all the test patterns as Bachian in nature. However, when a smaller amount of training data was used, the results were better and easier to understand in terms of the internal logic of the network. Training was done with pattern files created from melodic line only, melodic and rhythmic lines, intervals between adjacent melodic notes and frequency of occurrence of notes in a discrete sample.