Automatic discovery of parallelisms and hierarchy in music

Master's Thesis by Søren Tjagvad Madsen and Martin Elmer Jørgensen (2003)

The thesis is on musical similarities and grammatical structures of those. It can be downloaded here: gzipped post script (11.4 MB - a lot more unpacked!), pdf (3.4 MB - some reduction in the quality, the text looks blurred but is fine on print).

An introduction to our similarity finder system including source code and examples can be reached here: simfinder.tar.gz.


Abstract

There are three main ideas in this thesis.

Graph representation for non-monophonic music

We have attempted to create a datastructure for the symbolic representation of music. Based on the two most fundamental temporal relations found in music scores, precedence and simultaneity, we have developed a graph representation of music. A music graph may represent both monophonic and non-monophonic music. The music graph is a powerful and flexible abstraction that can be extended considerably to represent almost any symbol found in notated music scores. It is our attempt to break out of the sequential/parallel dichotomy used in so many representations in the literature, while still basing the representation on precedence and simultaneity. We argue that a graph representation is a more elegant representation, because it doesn't necessitate repeated restructuring of the data structure switching between the two fundamental modes (sequential or parallel structures) during the analysis and segmentation of music.

Search for musical parallelism

To search for similarities in music graphs, we have implemented the SimFinder system. The SimFinder uses a genetic algorithm to search for similarities in a music graph according to different similarity measures. Whereas much of the literature considers 'only' monophonic music, the SimFinder may search a music graph for either monophonic or non-monophonic similarities. We have experimented with several similarity measures for both monophonic and non-monophonic similarity. The SimFinder is a modular framework that allows new similarity measures to be defined and easily substituted for the ones we have defined. Similarity measures are built using a multiple viewpoint system that allows us to analyse music from a multitude of perspectives at once. By designing and mixing the right viewpoints we are able to specify what we are searching for. The SimFinder is designed so that many different kinds of viewpoints may be combined and tested experimentally.

Segmentation using a graph grammar

We describe a selection of past uses of formal grammars for the description of music. Just how powerful a grammar must be to be able to adequately describe music is a subject of debate, but it seems that something more than a context-free grammar is needed. We argue the importance of hierarchy in music and present a segmentation algorithm called the SimSegmenter for non-monophonic music, which is based on the SimFinder. The SimSegmenter builds a graph grammar to hierarchically describe a piece of music represented in a music graph. We present some test runs on fugue and chorale subjects by J.S.Bach.

Keywords

automated music analysis, non-monophonic music, multiple viewpoint systems, hierarchy and formal grammars, musical similarity and parallelism, graph representation of music, symbolic representations of music, graph grammar.