Structured Prediction for Piano Fingering with an Interval-Conditioned CRF

Automatic piano fingering is a structured prediction problem: the same notes admit many valid fingerings, expert annotators disagree, and a good choice at one note often depends on notes several positions away. We present KeyGenius, a compact (4.6M-parameter) Transformer encoder paired with a linear-chain CRF whose transitions are conditioned on local interval context, pretrained on a large weakly-labelled corpus and fine-tuned on the PIG dataset. On PIG’s canonical multi-annotator test set, the model is competitive with substantially larger ThumbSetpretrained autoregressive baselines — ahead on soft-match (Msoft 87.9 vs 86.8) and slightly behind the best on the primary metric (Mgen 65.8 vs 66.8) — at a small fraction of their parameter count. We further report three results that qualify common assumptions in the task: structured CRF decoding does not significantly improve aggregate annotator agreement over independent per-note prediction, though it more than halves runs of a single finger held across moving pitch; hand-authored ergonomic decoding constraints reduce rather than improve agreement with expert annotators; and pairwise inter-annotator agreement is a weak “ceiling” for a model predicting consensus. Finally, we describe a deployed system in active use by piano teachers. Code is openly available. Available at https://keygenius.app/