Change is Key: A Generative Framework for Controllable Musical Modulations

The ability to transition between two tonalities, or to modulate, is a vital compositional tool for shaping emotional contour and musical tension. However, controlling the generation of key changes in a meaningful and musically coherent manner is a significant challenge for AI-driven music systems, owing to the lack of large scale annotated modulation datasets. To overcome this data scarcity problem, we formulate musical modulation as a generative infilling task that enables models to learn from vast amounts of unannotated data via self-supervised training. We evaluate this framework through a purpose-built LLaMA-based infilling model which supports both an autonomous generation mode and a rule-guided model conditioned on harmonic progressions. To assess the modulatory mechanics without relying on ground-truth annotations, we design a multi-dimensional, reference-free evaluation suite quantifying stylistic coherence, rhythmic continuity, voice-leading smoothness, and harmonic tension. We validate these metrics through a subjective listening test to demonstrate a meaningful correlation with human aesthetic perception. Furthermore, zero-shot evaluations on both human-composed reconstructions and forced modulations across varying semitone shifts show that our autonomous model mitigates harmonic shock and maintains stylistic cohesion, even when navigating distant tonal intervals.