Abstract: Music is plentiful, but labeled data for music theory tasks like roman numeral analysis is scarce. Self-supervised pretraining on unlabeled data is therefore a promising means of improving performance on these tasks, especially because, during pretraining, a model may be expected to acquire latent representations of musical abstractions like keys and chords. However, existing deep learning models for roman numeral analysis have not used pretraining, instead training from scratch on labeled data, while conversely, pretrained models for music understanding have generally been applied to sequence-level tasks not involving explicit music theory, like composer or genre classification. In contrast, this paper applies pretraining methods to a music theory task by fine-tuning a masked language model, MusicBERT, for roman numeral analysis. We apply token classification to predict labels for each note, then aggregate the predictions of simultaneous notes to obtain a single label at each time step. Conditioning the chord predictions on key predictions gives more coherent labels. The resulting model outperforms previous roman numeral analysis models by a substantial margin.

Abstract: This article borrows from William Caplin's theory of formal functions, adapting these concepts to music from Bach's keyboard dance suites. It begins with a description of how Caplinian phrase functions such as "initiation" and "continuation" can be applied to this music. A novel "reinitiation" function is then proposed to identify a phrase function that Bach uses following non-final cadences. Reinitiation is similar to Caplin's initiating phrase functions, with the crucial difference that it occurs in an unstable harmonic con text. This combination of harmonic instability with otherwise initiating features allows reinitiation to succinctly express both a local "beginning" and a higher-level "middle." Bach most often achieves the harmonic instability characteristic of reinitiation function by means of a technique this article calls "de-tonicization": the immediate tonal desta bilization of a point of cadential arrival by reinterpreting it as a non-tonic harmony in another key. These ideas help illuminate how each moment of Bach's music "expresses [its] own location within musical time" (Caplin 2010).

Abstract: ‘When in Rome’ brings together all human-made, computer-encoded, functional harmonic analyses of music. This amounts in total to over 2,000 analyses of 1,500 distinct works. The most obvious motivation is scale: gathering these datasets together leads to a corpus large and varied enough for tasks including machine learning for automatic analysis, composition, and classification, as well as at-scale anthology creation and more. Further benefits include bringing together a range of different composers and genres (previous datasets typically limit themselves to one context), and of analytical perspectives on those works. We offer this data in as ready-to-use and reproducible a state as possible at http://github.com/MarkGotham/When-in-Rome, with code and documentation for all tasks reported here, including corpus conversion routines and feature extraction.

Abstract: The topic of this thesis is the properties of scales and modes other than the familiar major and minor, and the possibilities these afford for extended tonality and alternative functional harmony. In chapter one, I develop a systematic account of how to describe a large set of musically useful scales by using a generated line of intervals. This account is based on Hook's (2011) use of the line of fifths, but generalized to any eligible generating interval. In chapter two, I review the theoretical literature on various relevant scale properties, such as transpositional asymmetry and intervallic content. I discuss the implications of these properties when modes of these scales are realized in actual music. In chapter three, I discuss the placement of "tendency tones"—minor seconds that resolve to a member of the tonic triad—in various scales and modes, and how this placement affects the tonal implications of a given mode. In connection with this argument I introduce the idea of modal pitch-class sums and show that a mapping exists between these sums and a scale's structure upon the line of fifths. Finally, chapter four discusses how different structures of harmonic functions (analogous to the familiar tonic, subdominant, and dominant functions) can be realized within a mode. Building on the account of tendency tones developed previously, I discuss the contribution of these functions to the tonal dynamics of a given mode.