Repetition sans Ennui: Human and Algorithmic Discovery of Patterns in Music

Abstract

Finding patterns is a common act in human intellectual endeavours, and it is a complex challenge tackled by both humans and algorithms. For several decades, musical pattern discovery algorithms have been researched, and researchers have been comparing human-annotated patterns to algorithmic outputs, as well as algorithms to algorithms. However, traditional metrics have not fully captured the rich insights that these comparisons could offer. To contribute to the comparisons between musical pattern discovery mechanisms, this dissertation spans seven chapters.

Chapter 1 provides the background of the dissertation, including an overview, research approaches, contexts, scope, thesis statement, and an enumeration of contributions. Chapter 2 delves into the concept of musical patterns and explores the diverse landscape of musical pattern discovery algorithms. Our exploration reveals the complexities surrounding the definition of patterns and the multifaceted nature of these algorithms. Chapter 3 is dedicated to the collection tools for human-annotated musical patterns and the analysis of factors that influence annotations. We observe that musical background impacts annotated patterns; tool interfaces and automatic matching affect the length and frequency of annotations. Chapter 4 introduces four methods tailored for comparing musical pattern discovery algorithms. These methods provide novel insights into the discrepancies between human-annotated patterns and their algorithmically extracted counterparts. These methods provide a more comprehensive approach to comparing algorithms, aiding in the interpretation and evaluation of algorithmic outputs. Chapter 5 implements Pattrans, a Domain-Specific Language (DSL) in the functional language Haskell for comparing musical pattern occurrences through musical transformations. We delve into its design for uncovering the relations between pattern occurrences in a modular way. Chapter 6 employs Pattrans to scrutinise transformations between occurrences of musical patterns. Amongst other findings, we find that human-annotated patterns tend to have a higher proportion of exact repetitions and that different algorithms exhibit varying proportions of transformation compared to human annotations, contributing to a more nuanced view of pattern comparisons.

In summary, this dissertation not only contributes fresh perspectives to the comparison of musical patterns, but also introduces methods and tools that enrich the field of musical pattern discovery. We examine the concept of musical pattern, conduct pattern annotation experiments, and visualise and analyse human-annotated and algorithmically extracted patterns. In addition, we recognise the potential of the musical transformations that lie behind repeating and varying pattern occurrences. Using Haskell, we model the relationship between patterns and transformations. Following this, we investigate how to employ transformations to relate and classify musical pattern occurrences. Throughout our journey, we advocate for a more comprehensive approach to pattern comparison, extending beyond traditional metrics.