MAJEPPA

UMAP · 938 segments

Main MAJEPPA Space

A global view of held-out performance segments from the PerformanceMOS validation split, projected from MAJEPPA’s 1536-d backbone via UMAP. Recolor the cloud by performer skill, recording context, composer, or piece — and connect same-piece performances with edges to see how MAJEPPA pulls different renditions of the same score together.

Look for: skill clusters that emerge without any supervision, and the way performances of the same piece across beginner, intermediate, and virtuoso levels still cluster nearby.

PCA · 7 trajectories

Practice Journey

Seven longitudinal practice trajectories — each one piece, recorded by a single learner across many sessions or days. Sourced from YouTube practice playlists, with chronological order encoded as connecting lines and a white→blue color gradient marking early to latest sessions.

Note: each point is the session-averaged representation of an entire practice recording — the MAJEPPA backbone embeddings of all 300-token chunks within a session are mean-pooled into one vector per session.

Look for: trajectories that drift in a consistent direction as the learner improves. Click any point to play the audio for that day’s practice recording.

Piece-local UMAP · 25 études

Mistake Topology

Burgmüller Op.100 4-second performance windows, labeled as either annotated mistake regions or non-mistake regions via DTW alignment to the score. Each étude is projected separately (piece-local UMAP) so within-piece structure isn’t washed out by between-piece variance.

Look for: whether mistakes form a separable region from clean playing in the same piece. Click a point to hear both the performance window and the corresponding score segment side-by-side.