Modeling and Analyzing Musical Structure through Chord Progressions using Statistical and Machine Learning Methods

Abstract

This thesis investigates the modeling and analysis of musical structure through symbolic chord progressions. It integrates statistical, latent-space, and sequential modeling approaches to study harmonic relationships across genres, decades, and song sections. By utilizing the large-scale Chordonomicon dataset, this work examines both interpretable representations and predictive architectures, with emphasis on part-level segmentation and decomposed chord representations (root, quality+extensions, bass) for capturing fine-grained harmonic relationships. The methodological framework is split in two parts. The first pipeline applies statistical n-gram analysis and latent distance models to embed chord progressions into a continuous space, providing an interpretable visualization of musically meaningful patterns for genre and decade classification. To complement this approach sequential models are also employed to compare the results. The second pipeline adopts recurrent neural networks, their gated variants, and state-space models to predict the harmonic flow of chord sequences. This method assesses the impact of training set size, the length of the sequence available before prediction, chord decompositions and part-level context on performance. Descriptive analyses reveal strong harmonic regularities, stability of song structure across genres and decades, and systematic variations in chord transitions across song segments. For classification tasks, latent distance models provide interpretable representations that partially capture harmonic and stylistic relationships, while sequential architectures complement these findings by modeling temporal dependencies. Overall, the discriminative power across all models remains limited, reflecting the shared harmonic principles drive stylistic and temporal variations. This work demonstrates that integrating interpretable statistical modeling with efficient sequential architectures provides meaningful insights into the structure of symbolic music. The results highlight the significance of part-level context, chord decomposition, and harmonic representation in capturing the organizational principles of tonal music. These findings contribute to the field of music information retrieval and aid in understanding classification and predictive tasks, offering a foundation for future research in computational music analysis, stylistic characterization and automatic composition.