MAJEPPA: one AI model generates and assesses piano from beginner to virtuoso

MAJEPPA (Morphing and Assessing in a Unified Piano Performance Space) tackles a long-standing challenge: building one AI model that understands piano performances across the entire skill spectrum. Most prior systems either generate MIDI from scores or assess recordings, but not both. The team adapts a single pre-trained MIDI autoregressive transformer with a joint learning objective. Next-token prediction enables score-conditioned generation at various ability levels, while InfoNCE and supervised contrastive losses pull abstract score and performance representations into a shared embedding space. The result is a unified framework that can morph between performance styles and evaluate them consistently.

To train and test this, the authors curated the MAJEPPA dataset—about 4,000 annotated recordings spanning six expertise levels (from novice to virtuoso) and six recording contexts (e.g., practice room, studio, live concert). They also built EVPMR, a downstream benchmark suite covering quality assessment, competition ranking, and classification of mistakes and techniques. Experiments show that the learned representations transfer well across these tasks, suggesting a real-world path toward AI that can coach students, rank performances, and even generate stylistically appropriate practice material. This is a step toward unifying generation and understanding in music AI.