MAJEPPA

Training Data

MAJEPPA is trained on piano performances that span the full spectrum of skill and context. Each performance MIDI is paired with its musical score and annotated along two axes: performer expertise (child beginner, adult beginner, adult intermediate, child professional, piano teacher, virtuoso) and recording context (practice, sight-reading, showcase, demonstration, slow demonstration, concert).

Below are examples from each recording context.

Concert Performance

Professional recordings from the ATEPP dataset.

Showcase Recording

Different skill levels of performers from the MAJEPPA collection, recorded in informal settings.

Intermediate — Practice Session

Learners documenting their practice progress.

Sight-reading

Sight-reading sessions, capturing the early stages of musical development.

Demonstration

Demonstration classes from piano teachers.

Annotation Reliability & Quality Control

Each recording is manually labelled along two orthogonal axes — performer expertise (6 levels) and recording context (6 settings) — by conservatoire-trained student pianists. Because these labels are used both as conditioning tokens at training time and as the supervision signal for the contrastive condition loss L_cond, we treat label quality as a first-class concern.

• Annotator pool. Recordings were labelled by a small team of conservatoire student pianists (advanced undergraduate / graduate level), all of whom share a calibration document defining each category and showing borderline reference clips.
• Double labelling. A subset of recordings was independently labelled by two annotators to spot-check consistency; mismatches were resolved by a third reviewer. Inter-annotator agreement was measured on this subset using Cohen’s κ for each axis separately.
• Ambiguous cases. The most frequently confused boundary is adult beginner vs. adult intermediate. Annotators rely on multiple cues: tempo handling, hand independence on contrapuntal passages, expressive shaping, and the difficulty level of the chosen repertoire (e.g. RCM/ABRSM grade equivalent). When ambiguity could not be resolved, we conservatively defaulted to the lower level so the conditioning signal does not over-claim performer skill.
• Filtering. Recordings containing speech, metronome clicks, multi-instrument backing, or audio quality too poor for transcription were removed before labelling, so the labelled distribution reflects only material that survives the full pipeline.

Continuous Condition Similarity w_ij

The supervised contrastive condition loss L_cond in §4.3 of the paper does not treat conditions as discrete classes — it pulls pairs together by a continuous target similarity w_ij ∈ [0, 1] derived from the two label axes. The exact construction is:

1. Map each label to an integer ordinal score on each axis. The orderings reflect assumed proximity in performance behaviour, not a quality ranking:

Two design choices are worth flagging. Child beginner and adult beginner collapse to the same skill score (0): they share a comparably limited technical vocabulary, even if their typical repertoire differs. The three teacher-led demonstration settings (slow-demo, demo, fast-demo) collapse to the same formality score (2): they differ mainly in tempo, which the audio-to-MIDI tokenisation already encodes.

2. Compute the average normalised ordinal distance between two recordings i and j, then convert it to a similarity:

Each axis is normalised by its own range (max ordinal distance = 4), then the two normalised distances are averaged. The two axes are therefore weighted equally: a one-step gap on the skill axis contributes the same to w_ij as a one-step gap on the recording-context axis. The factor of 2 in the denominator keeps w_ij in [0, 1], with 1.0 reserved for pairs that share the same ordinal scores on both axes.

At training time, L_cond is computed only over cross-condition pairs — pairs that share the same skill and the same formality are masked out, since same-condition behaviour is handled by the within-condition objective L_self. This means w_ij only ever takes effect when at least one axis differs. The result reported on Pianist8 in §5.1 is sensitive to this design: collapsing the two beginner classes, equal-weighting the two axes, and using ordinal rather than one-hot targets each shift the implicit class structure that L_cond imposes on the embedding space.

Possible Dataset Confounds

Because we operate on transcribed MIDI rather than raw audio, the most common audio-side confounds — instrument timbre, room acoustics, microphone choice, upload codec, ambient noise — are stripped out before tokenisation. The representations therefore cannot directly exploit channel cues such as concert-hall reverb or webcam lo-fi.

However, the transcription step itself is not confound-free. Concert recordings are studio-clean and sit close to the training distribution of the transcriber (Transkun), while practice recordings are often captured on phones in untreated rooms with pedal noise, room tone, and occasional speech bleeding through. This means transcription error rates almost certainly correlate with our recording-context label, and to a lesser extent with skill (lower-skilled performers are more often recorded in informal settings). A label boundary that we attribute to performance behaviour could in principle be partly explained by systematic differences in transcription noise — e.g. dropped soft notes in quiet practice recordings, or spurious onsets near pedal release artefacts. The pedal-token ablation (§5) is one diagnostic for this; we do not claim it eliminates the concern.

A second, milder concern is repertoire correlation: virtuoso recordings cluster on the virtuoso repertoire, and beginner recordings on graded study material. Since score features go through the same tokeniser as performance features, the model can in principle infer skill from what is being played, not just how. The conditioning tokens are designed to absorb this signal explicitly so the contrastive losses can focus on performance-side variation, but the residual correlation is not fully separable from the data we have.