Improving Probabilistic Models of Melody

• Main Author: Chiu, Matt
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2022
• Subjects: Music theory
• ISBN: 9798352906125

This dissertation explores and improves probabilistic modeling of melody. The project is framed with an introduction (Chapter 1) and future directions (Chapter 5) and is otherwise divided into two sections: 1) a skip-gram model for evaluating non-adjacent dependencies (Chapter 2) and 2) a Gestalt-based model for melodic prediction (Chapters 3–4). Chapter 2 shows that non-adjacent pitch predictions at metrically strong beats are less reliable for prediction in the Essen folksong collection, suggesting that schematic patterns are less useful for describing the structure of this repertoire. Chapters 3–4 improves Temperley’s (2008) model by adding step inertia and motive factors. The impact of inertia is primarily style dependent: inertia is more useful for predicting pitches in the folksong dataset than the popular dataset—instead, popular songs tend to oscillate between two pitches. The motive factor greatly improves prediction in the popular music corpus, implying that exact repetition is perhaps one of the most useful factors for predicting pitches in popular music.