Music AI Research
Polygence Research Project — Predicting Piano-Piece Difficulty from MIDI with Ordinal Transformer Heads
Can AI tell how hard a piano piece is?
Piano teachers and exam boards spend years learning to judge how difficult a piece is — but what if a computer could do it automatically? This research builds an AI system that reads a digital music file (MIDI) and predicts its difficulty grade on the same scale used by major examination boards like ABRSM, RCM, and Trinity.
The model looks at what makes music hard to play — fast passages, hand independence, complex chords, pedaling — and learns to grade pieces on a 0–10 scale. Trained on nearly 8,000 piano pieces, it achieves state-of-the-art accuracy, outperforming both traditional approaches and prior machine learning methods.
Along the way, the paper uncovers why a popular AI technique (CORAL) fails dramatically on this problem, and proposes a better alternative (CORN) that improves accuracy by 9.4%. The full technical details are in the paper below.
Key Contributions
- 1.A new way to represent music for AI — capturing not just the notes, but how hands move, how loud they play, and how many notes sound at once
- 2.State-of-the-art difficulty prediction on the largest piano-syllabus dataset available (7,888 pieces)
- 3.Discovery of why a widely-used AI ranking technique fails on music — and a fix that improves accuracy by 9.4%