Melodic similarity algorithms for scores – a comparative evaluation of contrasting approaches

This thesis is concerned with melodic similarity algorithms for musical scores. The performance of two contrasting approaches is explored and the chosen algorithms fine tuned and evaluated using human observations of similarity. The most relevant musical features for assessing melodic similarity are identified from music perception research. Two contrasting algorithmic approaches are selected – a geometric algorithm and the string-matching edit distance approach. A number of different versions of both algorithms are implemented to assess the success of the musical features used. The internal weights of the algorithms are fine-tuned using a testbed of melodies for which human judgements of similarity have been gathered. These melodies are extracted from a piece of music in Theme and Variation form. While focusing on perceptual accuracy of the human similarity judgements, the best performing algorithms are identified and discussed. The internal algorithm weights are verified using additional extracts from the set of Theme and Variations. The ability of the algorithms to successfully generalise to a broader range of music is explored using two further collections of melodies in contrasting musical styles for which human observations of similarity exist.