PianoCoRe: Combined and Refined Piano MIDI Dataset

Abstract: Symbolic music datasets with matched scores and performances are essential for many music information retrieval (MIR) tasks. Yet, existing resources often cover a narrow range of composers, lack performance variety, omit note-level alignments, or use inconsistent naming formats. This work presents PianoCoRe, a large-scale piano MIDI dataset that unifies and refines major open-source piano corpora. The dataset contains 250,046 performances of 5,625 pieces written by 483 composers, totaling 21,763 h of performed music. PianoCoRe is released in tiered subsets to support different applications: from large-scale analysis and pre-training (PianoCoRe-C and deduplicated PianoCoRe-B) to expressive performance modeling with note-level score alignment (PianoCoRe-A/A*). The note-aligned subset, PianoCoRe-A, provides the largest open-source collection of 157,207 performances aligned to 1,591 scores to date. In addition to the dataset, the contributions are: (1) a MIDI quality classifier for detecting corrupted and score-like transcriptions and (2) RAScoP, an alignment refinement pipeline that cleans temporal alignment errors and interpolates missing notes. The analysis shows that the refinement reduces temporal noise and eliminates tempo outliers. Moreover, an expressive performance rendering model trained on PianoCoRe demonstrates improved robustness to unseen pieces compared to models trained on raw or smaller datasets. PianoCoRe provides a ready-to-use foundation for the next generation of expressive piano performance research.

• The paper introduces PianoCoRe, a legally sustainable collection of 250,046 performances covering 5,625 public-domain pieces and 483 composers, including the 157,207-performance PianoCoRe-A aligned tier.
• The authors combine content-based deduplication with an 80M-parameter MIDI quality classifier that achieves 89.1% macro F1, addressing unreliable source labels and transcription errors.
• The RAScoP refinement pipeline improves temporal alignment quality and downstream performance modeling, although lower-quality alignments can reduce rendering correlations and interpolation may create artifacts.

Motivation and scope

Score–performance pairs are the substrate for expressive performance analysis, rendering, transcription evaluation, and self-supervised symbolic pre-training. Existing piano MIDI resources split into two regimes: high-fidelity recorded corpora (MAESTRO, (n)ASAP) that are small, and AMT-derived corpora (GiantMIDI-Piano, ATEPP, Aria-MIDI) that are large but noisy, unaligned, and metadata-inconsistent. Combining them naively risks information leakage across train/test splits because naming conventions differ. PianoCoRe addresses this by unifying six open-source corpora under a single verified directory structure with content-based matching, quality labels, and note-level alignments (2605.06627).

The headline statistics: 250,046 performances of 5,625 pieces by 483 composers, totaling 21,763 hours of music, with scores available for 75.3% of performances. The note-aligned tier, PianoCoRe-A, provides 157,207 performances aligned to 1,591 scores — an order of magnitude larger than any prior open-source note-aligned collection ((n)ASAP offers roughly a thousand). The dataset is restricted to works in the public domain in the European Union, which the author presents as a deliberate legal-sustainability choice rather than a completeness guarantee.

Dataset construction

Matching is performed at the MIDI level rather than via title-based entity resolution, on the grounds that mislabeled files make metadata unreliable. Scores from MusicXML are converted to MIDI with partitura after processing dynamics/tempo markings, unrolling ornaments and grace notes, and generating both maximal (all repeats unfolded) and minimal repeat variants. Candidate score–performance pairs are pre-filtered by composer name, note ratio $R_n \in [0.75, 1.33]$, and catalog/key keywords, then verified with Parangonar's DualDTWNoteMatcher — accelerated roughly 12× via Numba JIT compilation, which the author states was essential for millions of pairwise alignments. A pair is accepted if alignment recall $R_a > 0.7$.

Sources were individually repaired before merging: (n)ASAP score MIDI files regenerated to fix zero-duration notes and implausible time signatures; ATEPP metadata corrected for duplicate movements and misattributed performances; GiantMIDI-Piano reduced to 2,139 files that could be verified against other datasets' scores and performances (a substantial cull from the 7,236-file curated subset, reflecting pervasive mislabeling); PERiScoPe stripped of ASAP/ATEPP overlap; and Aria-MIDI filtered to files with composer metadata. A shared artifact — "infinite" notes caused by unmatched note-on/note-off events in open-source transcribers — was repaired algorithmically, affecting 5,501 Aria-MIDI files alone.

Quality assessment and deduplication

PianoCoRe-B is derived from PianoCoRe-C through two mechanisms. First, content-based deduplication clusters performances sharing ≥50% of notes within a 50 ms onset tolerance, flagging 34,452 near-duplicates. Second, a learned MIDI quality classifier replaces unreliable source labels: the author shows that ATEPP's audio-derived "corrupted"/"low quality" annotations do not reliably reflect MIDI quality, with some well-aligned files labeled corrupted and clearly broken transcriptions left unlabeled.

The classifier is an 80M-parameter transformer encoder pre-trained with multi-mask language modeling on 371k Aria-MIDI files, fine-tuned on soft labels derived from alignment ratios ($R_a'$ thresholds), balanced with synthetic corruptions (note removal, onset/velocity jitter, insertions). It achieves a macro F1 of 89.1% on a held-out test set, perfectly separating score-like MIDI (F1 = 1.000). Notably, the NOMML heuristic from GigaMIDI flagged only 29 inexpressive performances in PianoCoRe — the author argues that onset jitter in transcriptions masks constant tempo, so heuristics designed for sequenced MIDI fail on transcribed data. This is a concrete, quantified argument for learned rather than heuristic filtering.

RAScoP alignment refinement

Raw aligner output suffers from temporal discontinuities (crossed links implying unrealistic tempi) and alignment holes (skipped repeats, transcription gaps). RAScoP processes alignments in four stages: hole detection via sliding-window sparsity ($H_w = 31$, $H_r = 0.75$), onset cleaning (removing intra-onset outliers beyond two standard deviations and correcting inter-onset intervals implying tempo outside 15–480 BPM), linear interpolation of unperformed notes with distance-weighted articulation/dynamics, and optional beat synchronization (not applied in the release, to preserve original timing).

The quantitative trade-off is modest: mean alignment recall drops from 0.935 to 0.920 after cleaning, with the share of sequences above 0.95 recall falling from 54.3% to 42.9% — but these migrate to lower bands rather than being discarded, and 86.6% retain recall above 0.85. The pipeline does not rematch aligner links; it only filters and interpolates, and default parameters were chosen empirically since automated tuning would require ground-truth human annotations.

Downstream validation

PianoFlow, a conditional flow-matching rendering model, was trained on progressively larger subsets. Three findings stand out:

• Scale delays overfitting: the ASAP-only model overfits quickly; adding ATEPP and PERiScoPe delays it, and the full PianoCoRe-A model had not converged by the end of training.
• Refinement matters: the model trained without RAScoP cleaning shows consistently higher validation loss on timing features, confirming that temporal noise removal improves learning.
• Robustness to unseen sources: in performance continuation, the full-dataset model achieves the best or second-best errors across all four test domains (ASAP, ATEPP, PERiScoPe, Aria-MIDI), e.g., velocity MAE of 7.872 bins on unseen Aria-MIDI versus 9.957 for the ASAP-only baseline.

One result cuts against a simple scaling narrative: including lower-quality alignments ($R_{\text{RAScoP}} \ge 0.7$) slightly hurts unconditional rendering correlations relative to the $\ge 0.85$ subset, indicating that interpolated-note density has a cost. Also, the ASAP+ATEPP model correlates better with the average PianoCoRe-A performance than the full model does, which the author attributes to ATEPP's concentration on renowned pianists — a distributional effect worth noting when interpreting the robustness gains.

Limitations

The author concedes several points plainly. The repertoire is skewed toward Western classical music and popular composers, inheriting source-corpus bias. Score errors in MusicXML can propagate downstream, and fewer than 1% of performances may contain neighboring movements or deviate substantially from their scores — composition-wise splits are recommended. Quality labels are calibrated for recall on corrupted/score-like classes and do not guarantee agreement with human judgment. RAScoP's interpolation can introduce deadpan segments and does not handle sustain pedal effects; the author suggests model-based prediction of missing notes and pedals as the proper remedy but leaves this open. The stated error margin of ~1% for movement-level naming errors inherited from sources is an assumption rather than a measured bound. Finally, no listening test was conducted, so whether objective metric improvements translate to perceptual quality remains an open question.

Conclusion

PianoCoRe consolidates fragmented piano MIDI resources into a tiered, legally constrained, metadata-verified collection whose aligned subset is unprecedented in scale. Its two technical contributions — a score-agnostic MIDI quality classifier and the RAScoP refinement pipeline — address failure modes (unreliable source labels, noisy alignments) that prior work handled ad hoc or not at all, and both are released as reusable tools. The downstream experiments support, though do not conclusively prove, the claim that combined refined data yields more robust performance models; the residual questions concern perceptual validation, source-balanced sampling given Aria-MIDI dominance, and pedal-aware interpolation.