PianoCoRe: Combined and Refined Piano MIDI Dataset

Ilya Borovik

doi:10.5334/tismir.333

PianoCoRe: Combined and Refined Piano MIDI Dataset

Transactions of the International Society for Music Information Retrieval

Volume 9 (2026): Issue 1

By: Ilya Borovik

Open Access

|Apr 2026

Figures & Tables

Table 1

Comparison of major symbolic piano performance datasets and PianoCoRe dataset with its tiers. Sources: R—recorded (Disklavier/Hardware), T—transcribed (Audio‑to‑MIDI), T‑HQ—transcribed labeled as high quality. Metadata: P—performer, S—piano solo probability, D—deduplication flag, Q—quality label. ^†Annotations are not available for all performances. ^‡Number of unique composer names computed from raw metadata, not manually verified.

Dataset	Composers	Pieces	Performances	Hours	Sources	Scores	Alignments	Metadata
MAESTRO	43	–	1,276	199	R	no	no	P
(n)ASAP	16	222	1,067	92	R	100%	beat/note	P
GiantMIDI	2,786	10,855	10,855	1,237	T	no	no	S
ATEPP	25	1,596	11,742	1,009	T	43.6%	no	P, Q^†
Aria‑MIDI	19,021^‡	–	1,186,253	100,629	T	no	no	S, P^†
PERiScoPe	82	2,738	46,473	3,784	R, T	81.9%	note	P^†
PianoCoRe‑C	483	5,625	250,046	21,763	R, T	75.3%	no	P^†
PianoCoRe‑B	478	5,591	214,092	18,757	R, T	75.0%	no	P^†, D, Q
PianoCoRe‑A	151	1,591	157,207	12,509	R, T	100%	note	P^†, D, Q
PianoCoRe‑A*	137	1,517	130,275	10,330	R, T‑HQ	100%	note	P^†, D, Q

The three‑stage data matching and annotation pipeline used to create PianoCoRe dataset.

Statistical overview of the **PianoCoRe‑C** dataset for the 50 most represented composers. **Top:** The total number of unique pieces per composer (blue) and the number of pieces with a musical score (light blue). **Bottom:** The average number of performances per piece, accumulated by the MIDI source.

Distribution of the number of musical pieces by the number of performances in PianoCoRe‑C.

MIDI performances from ASAP (orange) and ATEPP (blue) grouped by original labels and mapped as a function of performance‑to‑score note ratio $R_{n}$ and adjusted alignment ratio $R_{a}^{'}$ .

Table 2

Distribution of MIDI quality labels computed using the alignment‑based heuristics for the deduplicated, aligned performances in PianoCoRe‑B.

HQ	LQ	C	No Label
170,312	4,545	140	40,597

Table 3

MIDI quality classification dataset splits.

	S	HQ	LQ	C
training	2,500	2,500	2,500	2,500
real	953	2,500	1,000	86
synth	1,547	0	1,500	2,414
test	200	200	200	54
calibration	662	6,525	893	54

Table 4

Evaluation of MIDI quality classifiers using F1 scores. Best scores in bold. no synth—no synthetic training data, mean—mean pooling (no [CLS]), no TL—no transformer layer before the classifier head, no MLM—token embeddings and classifier only. The last block shows feature‑masking ablations.

Model	S	HQ	LQ	C	Avg.
base	1.000	0.839	0.777	0.946	0.891
no synth	1.000	0.759	0.778	0.946	0.871
mean	1.000	0.828	0.752	0.881	0.865
mean, no TL	0.993	0.802	0.713	0.851	0.840
no MLM	0.995	0.773	0.667	0.842	0.819
mask Pitch	1.000	0.803	0.723	0.913	0.860
mask Timing	0.990	0.788	0.747	0.851	0.844
mask Velocity	1.000	0.834	0.776	0.893	0.876

Table 5

PianoCoRe dataset and its source subsets labeled by the MIDI quality classifier.

Source	S	HQ	LQ	C
ASAP	0	1,066	0	0
ATEPP	0	10,231	900	433
GiantMIDI	11	2,071	52	5
PERiScoPe	82	34,596	91	4
Aria‑MIDI	1,151	180,977	18,359	17
PianoCoRe	1,244	228,941	19,402	459

Real‑world alignment challenges motivating the RAScoP pipeline. **Top:** local timing errors (crossed links) and missing/extra notes. **Bottom:** large structural deviation from a missing score segment, causing incorrect links. Other performed notes remain usable. Alignments were computed with Parangonar.

Note‑level alignment and the RAScoP pipeline for alignment refinement. The processing steps are demonstrated using an artificial example containing all types of errors. Score notes are drawn in black and performance notes are drawn in blue and green.

Distribution of inter‑onset deviations and beat tempos for alignments before processing (‑), after hole processing (H), after onset cleaning (O), and after both hole and onset cleaning (H + O).

Table 6

Mean alignment recall $\overset{―}{R}$ after different alignment refinement stages and the ratio of sequences (%) inside different recall bands.

	Raw		After H		After H+O
Band	${\overset{―}{R}}_{A}$	%	${\overset{―}{R}}_{H}$	%	${\overset{―}{R}}_{H + O}$	%
0.95–1.00	0.975	54.3	0.975	53.9	0.973	42.9
0.90–0.95	0.929	26.6	0.929	26.7	0.928	30.4
0.85–0.90	0.879	10.1	0.878	10.0	0.878	13.3
0.80–0.85	0.828	4.7	0.828	4.6	0.828	6.5
0.75–0.80	0.779	2.1	0.778	2.2	0.777	3.2
0.70–0.75	0.725	1.1	0.727	1.0	0.728	1.6
0.60–0.70	0.660	0.7	0.663	1.1	0.661	1.5
0.00–0.60	0.471	0.4	0.464	0.5	0.462	0.6
all	0.935	100.0	0.934	100.0	0.920	100.0

Validation loss curves for PianoFlow trained on different subsets of the data. Larger and refined training datasets reduce overfitting in the long run.

Table 7

Correlation between the features of the rendered and PianoCoRe‑A performances. First row—intra‑set correlations, other rows—models trained on different data subsets. Vel—velocity, IOI—inter‑onset‑interval, OD—relative onset deviation, Art—sustained articulation. The best scores are in bold.

	Vel	IOI	OD	Art
Dataset	0.57±0.19	0.90±0.06	0.22±0.17	0.44±0.19
ASAP	0.37±0.17	0.83±0.11	0.07±0.15	0.28±0.13
+ ATEPP	0.42±0.16	0.85±0.11	0.12±0.14	0.35±0.15
+ PERiScoPe	0.41±0.17	0.86±0.11	0.11±0.17	0.36±0.17
PianoCoRe‑A	0.40±0.17	0.86±0.11	0.10±0.17	0.35±0.17
$R_{RAScoP} \geq 0.7$	0.39±0.16	0.85±0.11	0.09±0.16	0.35±0.18
w/o RAScoP	0.41±0.16	0.85±0.11	0.09±0.16	0.36±0.18

Table 8

Conditional performance rendering (performance continuation) results across training subsets and unseen source sequences. Size denotes the training set size. Vel—Velocity (MIDI bins), TS—TimeShift (s), TD—TimeDurationSustain (s). Lower is better; best values are in bold.

		ASAP			ATEPP			PERiScoPe			Aria‑MIDI
Dataset	Size	Vel	TS	TD	Vel	TS	TD	Vel	TS	TD	Vel	TS	TD
ASAP	1 k	9.885	0.023	0.187	9.928	0.022	0.206	9.893	0.023	0.230	9.957	0.027	0.275
+ ATEPP	6 k	9.157	0.017	0.168	8.230	0.015	0.191	8.782	0.016	0.216	8.721	0.019	0.252
+ PERiScoPe	25 k	8.851	0.016	0.154	7.888	0.013	0.189	8.117	0.015	0.192	8.133	0.017	0.230
PianoCoRe‑A	124 k	8.613	0.016	0.155	7.967	0.014	0.194	8.094	0.015	0.194	7.872	0.017	0.205
$R_{RAScoP} \geq 0.7$	141 k	8.631	0.016	0.158	7.944	0.014	0.196	8.071	0.015	0.194	7.921	0.017	0.206
w/o RAScoP	124 k	8.734	0.017	0.159	8.059	0.015	0.193	8.199	0.016	0.196	8.055	0.018	0.211