This work investigates whether an explainable harmonic decision model built for individual songs can be applied unchanged to independently gathered human annotations. We selected 19 excerpts from HookTheory that contain aligned melody and chord annotations, yielding 439 evaluable harmonic segments, and evaluated a frozen Explainable Melody Harmonizer (EMH). EMH generates chord sequences from melody alone using explicit criteria: melodic compatibility, harmonic function, persistence, cadence, meter‑aware segmentation, and a Viterbi‑like global optimization. Human chords are hidden during generation and used only for post‑hoc comparison. Evaluation results show exact chord agreement of 20.50%, chord‑root agreement of 31.66%, and harmonic‑function agreement of 47.62% (functions assessed relative to the key inferred by EMH). Decision‑level analysis reveals that 203 segments (46.24%) contain human chords outside the model’s candidate vocabulary, while 146 segments (33.26%) have chords present in the vocabulary but not selected. These findings separate representational limits from selection differences, avoiding the assumption that human references are the sole musically valid solutions. The study demonstrates a reproducible, interpretable approach for corpus‑based evaluation of harmonic decision‑making.
Review