HarmonyForge: A Deterministic Co-Creative System for Symbolic Music Arrangement
Dulf Vincent Genis, Shivam Patel, Muawiz Khan, Yiren Liu, Jeff Ginger, Yun Huang
Keywords: Symbolic Music Arrangement ; Deterministic Constraint Optimization ; Explainable AI ; Human-AI Co-Creation ; Voice-Leading Constraints
Presented In-person
Chamber ensembles often face participation barriers when scores lack parts matching their specific instrumentation. While generative AI models draft arrangements quickly, they introduce structural drift and formatting errors that force musicians into tedious repair tasks. We present HarmonyForge, an interactive co-creative system that separates musical drafting from explanation. A deterministic SATB (Soprano, Alto, Tenor, Bass) Logic Core generates rule-constrained, instrument-mapped MusicXML harmony parts for an uploaded melody, which users then revise in a note-level editor. Hard constraints ban illegal spacing, voice-crossing, and parallel fifths or octaves; a motion heuristic prefers small pitch-to-pitch changes. An on-demand conversational LLM Theory Inspector explains voice-leading rules in plain language and may suggest repairs, but it never writes to the score without user confirmation. The proposed demonstration lets attendees configure custom ensembles, generate deterministic arrangements, override note placements, trigger theory violations on edit, query the pedagogical assistant, and export MusicXML, MIDI, WAV, or PDF. HarmonyForge provides an open sandbox demonstrating how deterministic constraints and glass-box explainability preserve creative ownership.