Modeling 53-TET Chord Progressions with a Microtonal GPT-2

This work explores the generation of harmonic chord progressions in 53-tone equal temperament (53-TET) using a GPT architecture, conditioned on microtonal transformation type, musical style, formal structure, and tonality. The training corpus originates from iRealPro lead sheets parsed into symbolic chord sequences and rendered to MIDI-MPE with voicings. Data augmentation transposes each song through all 53 pitch classes and applies 14 modal transformations, each defined by a holdrian-comma interval vector that redistributes diatonic steps across the 53-TET continuum, producing distinct microtonal modes from seven base scale types. The resulting corpus comprises 672,840 paired files combining symbolic metadata with MIDI-MPE. Each chord is represented as two consecutive blocks within a single flat autoregressive stream sharing one vocabulary and one prediction head: a symbolic block (L1) encoding the chord label, duration, and surrounding form, and a realization block (L2) encoding its voicing as pitch–velocity tokens. We train and compare two variants that differ only in how L2 pitches are encoded: Model A emits MIDI notes with per-note MPE pitch-bend deviations that realize the 53-TET target directly as playable MIDI, while Model B emits pure 53-TET step indices over a holdrian-comma vocabulary and defers the translation to MIDI with pitch bend to a deterministic post-processing step at generation time. A four-dimensional psychoacoustic dissonance vector, characterizing each tetrad by the intervallic qualities of its third, fifth, and seventh together with an aggregate sensory dissonance value, is injected as a continuous positional embedding shared across all tokens of the chord.