PaperlayerView exact revision on arXiv

Source-linked AI summary

PianoCoRe: Combined and Refined Piano MIDI Dataset

Ilya Borovik

arXiv:2605.06627v1cs.SDcs.LG

TL;DR

Existing piano MIDI resources have limited coverage, inconsistent metadata, and incomplete quality and note-level alignment. PianoCoRe unifies and refines open-source corpora, and experiments show refined subsets improve expressive rendering through greater repertoire diversity and cleaner note features.

  • Problem

    Existing piano MIDI datasets often lack standardized metadata, broad coverage, reliable quality, and note-level score-performance alignments needed for MIR and expressive modeling.

  • Method

    PianoCoRe combines and refines open-source piano corpora using quality classification, deduplication, and RAScoP alignment refinement.

  • Results

    Refined PianoCoRe subsets improve expressive rendering through increased repertoire diversity and cleaner note features.

  • Takeaways & Limitations

    PianoCoRe supports reproducible research across performance analysis, pre-training, expressive rendering, and score-to-performance translation.

  • Takeaways & Limitations

    PianoCoRe remains skewed toward Western classical repertoire and popular composers, reflecting biases in its underlying open-source corpora.

Abstract

from arXiv · show

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.

1. Introduction

PianoCoRe addresses the need for symbolic piano data linking written scores with expressive performances by combining and refining major open-source piano corpora. It provides a large, tiered collection designed for performance analysis, modeling, and large-scale pre-training.

  • Motivation: Scores represent written compositions, while performances capture interpretive variations in timing, dynamics, and articulation relevant to MIR.Modeling their relationship supports analysis of how performers convey musical structure and emotion.
  • Contribution: PianoCoRe combines and refines the largest open-source piano corpora of scores and performances into one comprehensive dataset.The work unifies existing piano MIDI resources with verified metadata.
  • Dataset: 21,763 h of piano music span 250,046 performances of 5,625 pieces by 483 composers, with scores available for 75.3% of performances.These figures describe the scale and score coverage of the collection.
  • Dataset: PianoCoRe is released as PianoCoRe-C, PianoCoRe-B, PianoCoRe-A, and PianoCoRe-A* for mixed-source collection, pre-training, note-aligned modeling, and high-quality alignments.The subsets respectively provide complete data, deduplicated quality-assessed data, score-aligned performances, and the best-quality performances and alignments.
  • Contribution: The work contributes corpus matching, deduplication and alignment-based MIDI quality labeling, a trained filtering classifier, and RAScoP alignment refinement.These contributions support the curated PianoCoRe-B dataset and refinement of score-performance note alignments.

2. Related Work

Related piano MIDI datasets span high-fidelity human recordings and large-scale audio-to-MIDI transcriptions, but differ substantially in scale, alignment, metadata, and quality. These differences create challenges for combining datasets safely and consistently, motivating PianoCoRe’s unification and refinement.

  • Recorded performance datasets: Human-recorded MIDI datasets preserve expressive detail, with MAESTRO providing over 200 h of time-aligned audio-MIDI performances but modest scale and diversity.ASAP extends MAESTRO with scores and beat annotations, while (n)ASAP adds note-level alignments.
  • Recorded performance datasets: Specialized corpora provide detailed alignments or synchronized audio-MIDI data, including Batik-plays-Mozart, Vienna 4x22, and SMD’s 50 performances of 50 pieces by 11 composers.These datasets prioritize exceptional detail for focused analysis rather than broad coverage.
  • Transcription-based datasets: Audio-to-MIDI transcription avoids sensor-equipped piano recording, enabling much larger collections but often lacking scores, complete annotations, or reliable quality labels.GiantMIDI-Piano provides 1,237 h across 10,855 pieces, ATEPP provides over 1,007 h from 11,674 performances, and Aria-MIDI offers over 100,629 h.
  • Motivation for PianoCoRe: Combining these datasets directly risks information leakage because their structures and metadata conventions differ, while collections also face unverified metadata, legal, duplication, corruption, and score-audio limitations.These limitations motivate standardized organization, verified metadata, and quality refinement when constructing a unified corpus.
  • Multi-source datasets: Recent multi-source efforts combine existing corpora or add large-scale alignment, including GigaMIDI’s over 1.4 million MIDI files and PERiScoPe’s over 35,000 note-aligned score-performance pairs.GigaMIDI’s unnormalized piece titles complicate grouping, while PERiScoPe combines (n)ASAP and ATEPP with 2,158 h of web-collected transcribed audio.

3. PianoCoRe Dataset … 3.3 Source Performance Datasets

PianoCoRe is constructed by converting, matching, aligning, refining, and integrating heterogeneous piano MIDI scores and performances into a unified collection. Its pipeline combines content-based candidate filtering and note-level verification with source-specific cleaning and transcription-artifact correction.

  • 3.1 Notation and Definitions: PianoCoRe represents scores, performances, notes, and alignments using structured MIDI entities, while Recall, Precision, Note Ratio, and an adjusted alignment ratio characterize matching quality.The representation includes pitch, onset, duration, and velocity, and alignments permit missing score or performance notes.
  • 3.2 Data Matching Methodology: MIDI-to-MIDI matching compares musical content directly, including performances whose scores exist only as MIDI and performance-to-performance matches when scores are unavailable.This approach addresses mislabeled or inconsistently named files that metadata-based matching may miss.
  • 3.2.1 Score Processing: Score processing regenerates MusicXML-derived MIDI while embedding dynamics and tempo, expanding ornaments and grace notes, and correcting parsing issues such as unrealistic time signatures, cut measures, duplicates, and zero-duration notes.The same score-processing pipeline was applied to ASAP and ATEPP sources.
  • 3.2.2 Candidate Pair Selection: Candidate pairs are pre-filtered by matching composer names, requiring 0.75 ≤Rn ≤1.33, and checking catalog, key, and scale keywords before expensive alignment verification.The filtering step reduces the number of pairs requiring computationally intensive alignment.
  • 3.2.3 Note Alignment and Verification: 12 times faster alignment on ASAP enabled millions of pairwise comparisons, with definitive matches requiring Ra > 0.7.DualDTWNoteMatcher computes note-level alignments, and performances failing against maximal unfolded scores are matched to minimal ones to increase retention.
  • 3.3.6 Transcription Artifacts: Source integration included filtered PERiScoPe performances and common cleanup of transcription artifacts, including repairing notes with ‘infinite’ durations caused by unmatched note-on and note-off events.PERiScoPe contributed 34,773 remaining performances after excluding ASAP or ATEPP files.
  • 3.3.2 ATEPP: ATEPP matching assigned 39 ASAP scores to 827 performances, while metadata correction addressed 49 duplicate movements, 265 reassigned performances, 24 wrongly named movements, and 43 performances.Only 5,091 of 11,674 transcribed ATEPP performances were paired with scores without an alignment.
  • 3.3 Source Performance Datasets: GiantMIDI-Piano used 7,236 curated MIDI files but exposed duplicate YouTube IDs and composition-label inconsistencies; Aria-MIDI filtered 621,132 composer-attributed files from 1,186,253 transcriptions.The retained GiantMIDI-Piano subset contained 2,139 performances by 402 composers, while Aria-MIDI transcriptions did not distinguish pressed from sustained note durations.

3.4 Musical Score Data Sources · 3.5 Data Combination Process · 3.6 PianoCoRe-C Dataset

PianoCoRe-C was created by iteratively combining matched piano scores and performances from multiple public datasets with manual curation, standardized naming, and copyright filtering. The resulting raw, comprehensive collection contains 250,046 performances spanning 5,625 pieces and movements by 483 composers and supports large-scale analysis and pre-training.

  • 3.4 Musical Score Data Sources: Public-domain MusicXML scores from PDMX expanded the score library beyond ASAP and ATEPP, while KunstderFuge and ClassicalMIDI scores enriched annotated compositions.Copyrighted scores were not redistributed, and inexpressive solo-piano sequences were filtered using the GigaMIDI NOMML heuristic.
  • 3.5 Data Combination Process: The dataset was assembled through a semiautomated, iterative process combining data matching, note alignment, manual curation, and labeling to resolve ambiguities.The process gradually merged scores and performances into a structurally unified collection.
  • 3.5 Data Combination Process: ASAP and ATEPP formed the core structure, with ASAP performances and scores reorganized into ATEPP’s directory hierarchy and 21 additional ASAP pieces placed in new directories.This established a unified base of recorded and transcribed performances with corresponding scores.
  • 3.5 Data Combination Process: GiantMIDI-Piano, PERiScoPe, and Aria-MIDI performances were matched to available scores, with new score-and-performance directories added for previously absent pieces.Remaining performances were also matched against scoreless ATEPP performances and one another to increase coverage.
  • 3.5 Data Combination Process: Manual verification standardized composition and movement titles using IMSLP and web search, while inclusion was restricted to works public-domain in the European Union.The checks corrected mislabeled files and prevented duplicate cataloging under different names.
  • 3.6 PianoCoRe-C Dataset: 250,046 performance MIDI files in PianoCoRe-C cover 2,869 compositions and 5,625 pieces and movements by 483 composers across historical periods and styles.The dataset is described as the most diverse collection of piece-wise annotated piano performances.
  • 3.6 PianoCoRe-C Dataset: PianoCoRe-C is neither deduplicated nor quality-filtered, serving instead as the raw foundation for PianoCoRe-B and PianoCoRe-A.Its matched score and performance files support maximum-scale applications including self-supervised pre-training, large-scale music analysis, and data-cleaning research.

4. Performance MIDI Quality Assessment

This section refines PianoCoRe-C by removing near-duplicate performances and assigning direct MIDI-quality labels with a classifier trained to detect corrupted and score-like files. The resulting PianoCoRe-B subset contains 214,092 deduplicated performances suitable for piano-expression analysis.

  • Deduplication: 34,452 near-duplicates were flagged and removed from PianoCoRe-C, leaving lead and unique performances with duplicate status recorded in metadata.Performances were compared using close note-onset matches for notes sharing the same pitch, with a 50 ms onset-difference threshold.
  • Quality assessment: The MIDI-quality classifier directly assesses performances without requiring score or audio alignment, targeting corrupted transcriptions and score-like performances.This addresses the limitation that reference-based transcription metrics require ground-truth data, while onset jitter can make inexpressive score-like performances appear expressive to NOMML.
  • Quality assessment: Original ATEPP labels did not reliably reflect MIDI quality: some poorly aligned files labeled low quality or unlabeled were broken, while some corrupted files aligned well and remained usable.Recorded and high-quality performances generally had adjusted alignment ratios above 0.85, but corrupted files were scattered across alignment levels.
  • Quality assessment: 89.1% macro F1 was achieved by the best classifier configuration on the held-out test set, perfectly distinguishing score-like MIDI and reducing errors among HQ, LQ, and C.Synthetic training samples and token-based aggregation improved decision boundaries, while pitch, dynamic, and timing features jointly contributed to classification.
  • PianoCoRe-B: 214,092 deduplicated performance MIDI files remained in PianoCoRe-B after filtering out performances classified as Corrupted or Score, covering 5,591 pieces by 478 composers.PianoCoRe-B was obtained by applying both deduplication and quality-assessment models to PianoCoRe-C.

5. Refined Note Alignment

RAScoP refines raw score-performance alignments by removing structurally incorrect links, cleaning temporal errors, interpolating missing notes, and synchronizing performances to scores. Applied to PianoCoRe-B, it produces large cleaned aligned subsets, while reducing timing noise and tempo outliers at a modest recall cost.

  • Temporal refinement and interpolation: The pipeline removes large timing deviations and interpolates missing notes, producing note-level aligned MIDI while marking interpolations for filtering or model-training annotation.It also prevents duplicate pitch-onset notes and shortens overlapping notes at new key presses.
  • Pipeline overview: RAScoP transforms raw alignments into clean, complete, and temporally coherent score-performance pairs through hole processing, onset refinement, note interpolation, and synchronization.Each stage can be enabled independently, and the pipeline filters existing links rather than rematching them.
  • Hole processing: Alignment holes are detected as contiguous regions with sparse or nonsensical matches and removed to eliminate structurally incorrect score or performance segments.The default detector uses a 31-note window and a 0.75 unaligned-note threshold.
  • Effects of refinement: Applying the full H+O pipeline significantly reduces within-chord inter-onset deviation variability and stabilizes beat tempos around a musically plausible range.The refinement corrects extreme tempo values implied by raw noisy alignments.
  • Dataset outcomes: 1.5%: average alignment recall decreases from 0.935 to 0.920, while PianoCoRe-A contains 157,207 cleaned sequences spanning 1,591 pieces and 12,509 h.PianoCoRe-A* retains 130,275 performances for 1,517 pieces, requiring at least 85% aligned notes and High Quality MIDI.

6. Music Performance Rendering

PianoCoRe-A improves expressive piano performance rendering by combining larger, more diverse, and refined aligned data. Increasing dataset scale delays overfitting and generally improves robustness across unseen compositions and MIDI sources, while remaining limitations concern source imbalance and the lack of subjective validation.

  • Experimental setup: PianoFlow was trained on progressively larger PianoCoRe-A subsets after filtering performances below RRAScoP 0.85, with additional ablations using RRAScoP ≥0.7 and raw alignments.The experiments used ASAP, ASAP+ATEPP, ASAP+ATEPP+PERiScoPe, and the full dataset, splitting data by composition.
  • Validation: Larger and refined training subsets delay overfitting, whereas the model trained only on ASAP quickly overfits despite that dataset’s high quality.The validation losses were evaluated on unseen performances from the same source data.
  • Validation: Refined PianoCoRe-A produces more stable and consistently lower validation loss than w/o RAScoP, especially for note time shifts.The result indicates that removing temporal noise is important for learning accurate timing models.
  • Unconditional performance rendering: Models trained on more diverse datasets consistently outperform ASAP on multi-source unconditional rendering, while raw alignments reduce correlation for higher-quality timing features.More training data with interpolated notes slightly harms unconditional rendering, and the cleanup-free model performs worse on IOI and OD.
  • Performance continuation: PianoCoRe-A achieves the best average continuation performance on ASAP and Aria-MIDI and second-best results on the other datasets, supporting robustness across source distributions.Only the model without overrepresented Aria-MIDI achieves similar or lower errors on ATEPP and PERiScoPe, while the full-data model may improve with longer training.
  • Limitations and future work: Subjective listening tests and more balanced source sampling remain needed to assess perceptual gains and improve generalization across domains.Fine-tuning on high-fidelity subsets such as ASAP is suggested as a possible further improvement.

7. Limitations

PianoCoRe remains limited by inherited naming and repertoire biases, potentially erroneous MusicXML scores, and automated alignment and quality-labeling procedures. Some performances may contain movement or score mismatches, while RAScoP interpolation can create deadpan segments and does not model sustain pedals.

  • Dataset composition and naming: PianoCoRe reserves a 1% margin for inherited movement-level naming errors and remains skewed toward Western classical repertoire and popular composers.These limitations reflect biases and inconsistencies inherited from the underlying open-source corpora.
  • Source scores and alignment: Errors in open-source MusicXML scores may propagate downstream, while fewer than 1% of performances may contain neighboring movements or large score mismatches.The paper recommends composition-wise splits for applications using the dataset.
  • Quality labeling: Classifier-based MIDI quality labels prioritize recall for corrupted and score-like classes but do not guarantee alignment with human expectations.The labels were designed to filter incorrect and inexpressive data rather than ensure perfect perceptual agreement.
  • Alignment refinement: RAScoP note interpolation may introduce deadpan performance segments and does not handle sustain-pedal effects, requiring downstream correction.The paper suggests predicting missing notes and pedals as a better solution.

8. Conclusion

PianoCoRe unifies and refines open-source piano MIDI corpora into tiered subsets supporting diverse MIR and expressive performance tasks, while quality filtering and alignment refinement improve data integrity and model benefits. Future work targets broader repertoires, stronger quality assessment, and finer annotations.

  • Dataset and research foundation: PianoCoRe combines, refines, annotates, and aligns existing open-source piano MIDI corpora into tiered subsets for diverse research tasks.The subsets support performance analysis, large-scale pre-training, expressive piano performance rendering, and score-to-performance translation.
  • Data integrity: A MIDI quality classifier detects deadpan and corrupted transcriptions, while an alignment refinement pipeline removes temporal outliers from aligned score-performance data.These methods address performance-MIDI quality and note-level alignment integrity.
  • Data integrity: Models trained on refined subsets benefit from increased repertoire diversity and cleaner note features.The conclusion attributes these benefits to the refinement process and expanded data coverage.
  • Future directions: Future directions include multi-instrument repertoires, more robust quality assessment models, and more granular score and performance annotations.The paper aims for PianoCoRe to support continued symbolic music performance modeling and analysis research.

Ethical Statement

PianoCoRe was best-effort filtered under EU public-domain regulations, with composer metadata released transparently despite unavoidable accuracy limitations. The dataset and annotations use CC-BY-NC-SA 4.0, and no ethics approval or participant consent was required.

  • Copyright and intellectual property: PianoCoRe was best-effort filtered for works whose authors had been deceased for more than 70 years, but complete accuracy across thousands of diverse files was difficult.The criterion follows European Union public-domain regulations.
  • Copyright and intellectual property: Annotated composer metadata was released alongside the dataset to support transparency.
  • Licensing and ethics: The dataset, processed files, metadata, and alignment annotations were published under CC-BY-NC-SA 4.0, with no ethics approval or human participant consent required.The study processed publicly available MIDI data and involved no human subjects.

Author’s Contribution

Ilya Borovik led the paper’s conceptualization, methodology, software implementation, data curation, and manuscript writing.

  • Ilya Borovik handled the research conceptualization, methodology, software implementation, data curation, and manuscript writing.

Notes

The notes list links to sheet-music, classical MIDI, and copyright-related resources.

  • Notes: The cited resources include MuseScore, Kunst der Fuge, Classical MIDI, IMSLP, and an EU summary of copyright and related-rights protection terms.The passage provides URLs for each resource.
Loading 2605.06627v1…