Generative music models are usually evaluated on whether listeners like the output. That tells you very little about why a system chose a chord, and nothing at all about whether it could have chosen differently.
A paper posted to arXiv on 14 October 2026 takes a narrower and more useful question. Josef Pavlíček and Martin Pelikán ask whether an explainable harmonic decision model built for individual songs still works when pointed at independently collected human annotations — without changing its rules.
The setup
The model is the Explainable Melody Harmonizer (EMH), a rule-based system whose decisions can be inspected. The method matters more than the model:
- Freeze EMH. No retuning, no fitting to the new data.
- Run it on HookTheory excerpts that have aligned melody and chord annotations.
- Generate chords from the melody alone.
- Use the human chords only afterwards, for comparison.
That ordering is what makes the result interpretable. The model never sees the answer, so the comparison measures the model’s prior rather than its ability to fit.
The corpus is small and stated as such: 19 excerpts, 439 harmonic segments.
The three agreement rates
| Measure | Agreement |
|---|---|
| Exact chord | ~20.5% |
| Chord root | ~31.7% |
| Harmonic function | ~47.6% |
Read as a ladder, these are more informative than any one of them. The model picks the same chord a human picked one time in five; the same root roughly one time in three; and the same harmonic function — tonic, dominant, subdominant — nearly half the time.
The gap between 20.5% and 47.6% is where the musical content lives. A system agreeing on function but not on the specific chord is not making random choices; it is making defensible choices that differ in voicing, extension or substitution. Exact-match accuracy alone would have reported that as failure.
The decomposition that is the actual contribution
The headline rates are context. The paper’s real move is splitting the disagreements into two categories:
- In 46.24% of segments, the human chord was outside the model’s candidate set — the model could not have produced it at all.
- In 33.26% of segments, the human chord was available but not chosen.
These are different failures requiring different fixes. The first is a representational limit: the vocabulary, the rules, or the harmonic grammar cannot express what the annotator wrote. The second is a decision problem: the candidate was there and the ranking preferred something else.
Almost half the disagreement is representational. Tuning the model’s preferences would not move that half at all — the candidate generation would have to change first. Conversely, improving candidate generation without touching the ranking would leave the other third untouched.
Most evaluations of harmonisation systems collapse both into a single accuracy figure and so cannot tell a practitioner which of the two to work on.
What it does and does not establish
The paper presents this as a reproducible, interpretable way to evaluate harmonic decisions against a corpus rather than against a single song or a listening panel. The comments list 10 pages, 3 figures, 2 tables, and source code in git, so the procedure is inspectable.
The limits are real and worth stating. Nineteen excerpts is a small sample, HookTheory annotations are crowd-sourced and reflect one community’s analytical conventions, and a frozen rule-based harmoniser is not the same object as a trained neural model — the decomposition transfers as a method, not as a result about generative models generally.
The transferable idea is the two-bucket split. Any system that generates candidates and then ranks them — chord choice, drum pattern selection, colour palette suggestion, layout proposal — can be evaluated this way, and the answer tells you which half of the pipeline is responsible.
No journal reference is listed on the arXiv page.
Sources
- From Individual Songs to Music Corpora: Evaluating Explainable Harmonic Decision-Making on HookTheory — arXiv 2610.22416
- HookTheory
- arXiv listing — cs.SD