🤖 machine learning

PianoCoRe: Combined and Refined Piano MIDI Dataset

This paper introduces PianoCoRe, a large-scale, unified, and refined piano MIDI dataset comprising over 250,000 performances with note-level alignments, accompanied by novel tools for quality classification and alignment refinement to advance expressive performance modeling and music information retrieval research.

Original authors: Ilya Borovik

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ilya Borovik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to play the piano with human emotion. To do this, you need two things: the sheet music (the instructions) and a recording of a real person playing it (the expression).

For a long time, researchers had a problem: they had either a mountain of sheet music with no recordings, or thousands of recordings with no sheet music, or the two didn't match up correctly. It was like having a library of recipes but no photos of the finished dishes, or a million photos of dishes but no idea what ingredients were used.

This paper introduces PianoCoRe, a massive new "library" that fixes this mess. Here is how it works, broken down simply:

1. The Great Cleanup (The "Librarian" Work)

The author, Ilya Borovik, took several existing piles of piano data from around the internet and smashed them together. But instead of just dumping them in a heap, he acted like a super-organized librarian:

Matching: He used a smart computer program to match the sheet music to the correct recordings, even if the file names were messy or different.
Cleaning: He found and removed "duplicates" (like finding two copies of the same song played by the same person and keeping only one).
Quality Control: He built a "robot inspector" (a classifier) that looks at the recordings and says, "This one is garbage," "This one is just a computer reading the sheet music robotically," or "This one is a great human performance." He threw out the bad ones.

2. The "RAScoP" Pipeline (The "Editor")

Sometimes, the matching between the sheet music and the recording isn't perfect. Maybe the recording skipped a repeat sign, or the computer got confused and added a note that wasn't there.
The author created a tool called RAScoP (Refined Alignment for Scores and Performances). Think of this as a smart editor:

It finds the "glitches" where the timing is weird.
It fills in the gaps where the recording missed a note (by guessing what the note should have been based on the surrounding music).
It makes sure the sheet music and the recording are perfectly synchronized, note-for-note.

3. The Result: A Tiered Library

The final result is a huge dataset called PianoCoRe, which contains over 250,000 performances of 5,600 pieces by 483 composers. That's over 21,000 hours of music!

To make it useful for different people, the author organized it into "tiers" (like different sections in a library):

PianoCoRe-C (The Whole Collection): The massive, raw mix of everything. Good for big data projects.
PianoCoRe-B (The Clean Version): The duplicates and garbage are gone. This is the "safe" version for training AI models.
PianoCoRe-A (The Perfectly Aligned Version): The recordings that are perfectly matched to the sheet music, with the "editor" (RAScoP) fixing any timing errors. This is the gold standard for teaching robots to play expressively.

4. Did It Work? (The Proof)

The author tested this new library by training a robot to play piano.

The Test: He taught robots using small, messy datasets and compared them to a robot trained on his new, clean, huge dataset.
The Result: The robot trained on PianoCoRe was much better at playing with human-like timing and emotion. It didn't get confused as easily when playing new songs it had never seen before.

Why This Matters

Before this, researchers had to spend months cleaning up their own data, often ending up with small, messy datasets. PianoCoRe gives them a ready-to-use, high-quality foundation. It's like giving a chef a pre-chopped, pre-washed, perfectly measured box of ingredients instead of a sack of raw, muddy vegetables.

Important Note: The author was very careful to only include music that is in the "public domain" (old enough that copyright has expired), ensuring this library is legal and safe for everyone to use forever.

In short: PianoCoRe is the ultimate, cleaned-up, perfectly matched library of piano music that helps computers learn to play like humans.

Technical Summary: PianoCoRe: Combined and Refined Piano MIDI Dataset

1. Problem Statement

Symbolic music datasets containing matched scores and performances are critical for Music Information Retrieval (MIR) tasks, particularly for modeling expressive performance rendering. However, existing resources suffer from significant limitations:

Fragmentation: High-fidelity recorded datasets (e.g., MAESTRO, ASAP) offer precision but lack scale and stylistic diversity. Conversely, large-scale transcribed datasets (e.g., GiantMIDI-Piano, Aria-MIDI) offer scale but often contain transcription errors, lack precise note-level alignments, and suffer from inconsistent metadata.
Incompatibility: Incompatible naming schemes and metadata standards make combining datasets difficult without risking information leakage.
Quality Issues: Transcribed MIDI files often contain corrupted data, "score-like" inexpressive transcriptions, duplicates, and temporal noise (e.g., onset jitter, infinite pitch artifacts).
Legal Constraints: Many existing datasets include copyrighted material, limiting their long-term academic utility.

There is a critical gap for a unified, large-scale resource that combines the scale of transcribed data with the precision of recorded performances, all legally sustainable and aligned to scores.

2. Methodology

The paper presents PianoCoRe, a dataset constructed through a multi-stage curation and refinement pipeline.

2.1 Data Collection and Matching

The dataset integrates five primary open-source sources: ASAP, ATEPP, GiantMIDI-Piano, PERiScoPe, and Aria-MIDI.

Legal Filtering: Content is restricted to works in the public domain within the European Union to ensure legal sustainability.
Score Processing: MusicXML files are converted to MIDI using the partitura library. This involves unrolling ornaments (trills, mordents), expanding grace notes, and handling repeats (creating both maximal and minimal versions).
Matching Strategy: A three-stage pipeline matches performances to scores:
1. Candidate Selection: Filtering based on composer name, note count ratio ( $0.75 \le R_n \le 1.33$ ), and catalog keywords.
2. Note Alignment: Using DualDTWNoteMatcher (from Parangonar) with Numba-optimized Dynamic Time Warping (DTW) to compute note-level alignments.
3. Verification: A pair is accepted if the alignment recall ( $R_a$ ) exceeds 0.7.

2.2 Quality Assessment and Deduplication

To produce the PianoCoRe-B subset, the authors address duplicates and quality issues:

Deduplication: A content-based heuristic clusters performances with $\ge 50\%$ similar note onsets (within 50ms). One "lead" file is retained based on source priority.
MIDI Quality Classifier: A 12-layer transformer encoder (pre-trained on Aria-MIDI) is trained to classify MIDI files into four categories: Score (deadpan), High Quality (HQ), Low Quality (LQ), and Corrupted.
- Training Data: Labels are generated via a soft heuristic based on alignment ratios ( $R'_a$ ). Synthetic corruptions (note removal, jitter) are used to balance the dataset.
- Performance: The classifier achieves a macro F1 score of 89.1%, effectively distinguishing score-like transcriptions from expressive performances.

2.3 Alignment Refinement (RAScoP)

To create the PianoCoRe-A/A* subsets, the authors introduce RAScoP (Refined Alignment for Scores and Performances), a pipeline to clean raw alignments:

Alignment Hole Processing: Removes large regions of sparse or nonsensical alignments (e.g., unperformed repeats) using a sliding window approach.
Onset Cleaning: Filters intra-onset deviations (chord timing) and corrects inter-onset intervals that imply unrealistic tempos (15–480 BPM). It adjusts global timing to match local tempo estimates.
Note Interpolation: Linearly interpolates missing notes (deletions) based on neighboring performed notes, averaging duration and velocity. Interpolated notes are marked with MIDI text markers.
Synchronization: Aligns the beat structure of the performance to the score (optional in the final release to preserve original timing).

3. Key Contributions

PianoCoRe Dataset: A unified collection of 250,046 performances of 5,625 pieces by 483 composers, totaling 21,763 hours. It is released in tiered subsets:
- PianoCoRe-C: The complete, mixed-source collection.
- PianoCoRe-B: A deduplicated, quality-assessed subset (214,092 performances) for large-scale pre-training.
- PianoCoRe-A: A note-aligned subset (157,207 performances) with cleaned alignments.
- PianoCoRe-A:* A high-confidence subset (130,275 performances) with $\ge 85\%$ alignment recall.
MIDI Quality Classifier: A trained model to detect corrupted and score-like transcriptions without requiring ground-truth audio or scores.
RAScoP Pipeline: An open-source algorithm for refining note alignments, cleaning temporal outliers, and interpolating missing notes.
Legal Sustainability: The dataset is curated to comply with EU public domain regulations, ensuring long-term accessibility.

4. Results and Evaluation

The paper validates the dataset through an expressive performance rendering task using the PianoFlow model (conditional flow matching).

Training Convergence: Models trained on larger, refined datasets (PianoCoRe-A) showed delayed overfitting and lower validation losses compared to models trained on smaller subsets (e.g., ASAP only) or unrefined data.
Unconditional Generation: Models trained on PianoCoRe-A demonstrated improved robustness and higher Pearson correlations with ground-truth features (velocity, inter-onset intervals, articulation) on unseen test sets compared to baselines.
Performance Continuation: In a continuation task across four distinct source domains (ASAP, ATEPP, PERiScoPe, Aria-MIDI), the model trained on the full PianoCoRe-A dataset achieved the best or second-best performance across all domains, demonstrating robustness to varying data distributions.
Refinement Impact: The use of RAScoP (removing temporal noise) was shown to be crucial; models trained on raw alignments without cleaning performed worse on timing features (IOI and OD).

5. Significance and Claims

The paper claims that PianoCoRe provides a "ready-to-use foundation for the next generation of expressive piano performance research." Its significance lies in:

Scale and Diversity: It offers the largest open-source collection of note-aligned score-performance pairs to date (157,207 performances in PianoCoRe-A).
Quality Control: By integrating a quality classifier and the RAScoP pipeline, it addresses the "noise" problem inherent in large-scale transcribed datasets, enabling the training of more accurate and robust models.
Reproducibility: The tiered structure and unified metadata allow researchers to create non-overlapping data splits across previously isolated datasets, facilitating reproducible research.
Accessibility: The dataset is openly available on Zenodo and Hugging Face under a CC-BY-NC-SA 4.0 license, with code for the pipeline and classifier provided.

The authors note limitations, including a skew toward Western classical repertoire, potential residual errors in source MusicXML scores, and the fact that interpolation does not currently model sustain pedal effects. They suggest future work could extend these methodologies to multi-instrument repertoires and more granular annotations.