KRAISLER: A Multi‑Track Dataset of Piano and Violin Duet Recordings for Music Information Retrieval Research

Instruments	Dataset	Recorded Audio	Multi Track	Direct Piano MIDI Capture	Note Annotation	Beat Annotation
Solo piano	MAPS (Emiya et al., 2010)	✓		✓	✓
	SMD (Müller et al., 2011)	✓		✓	✓
	MAESTRO (Hawthorne et al., 2019)	✓		✓	✓
	ASAP (Foscarin et al., 2020)	✓		✓	✓	✓
Ensemble (no piano)	URMP (Li et al., 2018)	✓	✓		✓
Ensemble (no piano)	Bach10 (Duan and Pardo, 2011)	✓	✓		✓	✓
	PHENICX‑Anechoic (Miron et al., 2016)	✓	✓		✓
	ChoraleBricks (Balke et al., 2025)	✓	✓		✓
Piano Ensemble	RWC‑Classic (Goto et al., 2002)	✓			✓	✓
Piano Ensemble	TRIOS (Fritsch and Plumbley, 2013)	✓	✓		✓
	MusicNet (Thickstun et al., 2017)	✓			✓
	SWD (Weiß et al., 2021)	✓			✓
	PCD (Özer et al., 2023)	✓	✓		✓	✓
	KRAISLER (Ours)	✓	✓	✓	✓	✓

Table 2

The list of pieces in the KRAISLER dataset.

No.	Composer	Title	Key	Year	Duration
01	Tchaikovsky	Valse sentimentale, Op. 51 No. 6	B^♭ minor	1882	1:04
02	Rachmaninoff	Preghiera (arr. by Kreisler from Piano Concerto No. 2, Op. 18)	C minor	1940	1:03
03	Saint‑Saëns	Danse Macabre, Op. 40	G minor	1874	1:19
04	Tchaikovsky	Violin Concerto in D major, Op. 35, mov.1	D major	1878	1:18
05	Saint‑Saëns	Violin Concerto No. 3 in B minor, Op. 61, mov.1	B minor	1880	1:12
06	Schumann	3 Romances, Op. 94	A minor	1849	1:18
07	Brahms	Violin Sonata No. 1 in G major, Op. 78, mov.3	G major	1879	1:04
08	Grieg	Violin Sonata No. 3 in C minor, Op. 45, mov.1	C minor	1887	1:14
09	Franck	Violin Sonata in A major, mov.4	A major	1886	1:24
10	Chopin	Nocturne No. 20 in $C^{♯}$ minor, Op. Posth.	$C^{♯}$ minor	1830	1:23
11	Schumann	Dichterliebe, Op. 48, No. 1	$F^{♯}$ minor	1840	1:24
12	Schumann	Dichterliebe, Op. 48, No. 5	A major	1840	0:48
13	Beethoven	Violin Sonata No. 5 in F major, Op. 24, mov.1	F major	1801	0:50
14	Ponce	Estrellita (arr. by Heifetz)	$F^{♯}$ major	1912	1:17
15	Rachmaninoff	Vocalise, Op. 34, No. 14	$C^{♯}$ minor	1912	1:54
16	Tchaikovsky	Mélodie from Souvenir d’un Lieu Cher, Op. 42, No. 3	E^♭ major	1878	1:28
17	Mozart	Violin Concerto No. 3 in G major, K. 216, mov.1	G major	1775	1:19
18	Kreisler	Liebesleid (Love’s sorrow)	F minor	1905	1:25
19	Kreisler	Liebesfreud (Love’s joy)	D major	1905	1:07
20	Mendelssohn	Violin Concerto in E minor, Op. 64, mov.1	E minor	1844	1:09
Total duration					25:00

Recording setup for the dataset: **(a)** Room 1 contains the Disklavier piano, while **(b)** Room 2 is used for violin recording, with a soundproof window enabling visual interaction between performers.

Table 3

Dataset components for each excerpt.

Data Types	Components	Format
Audio (dry/studio/hall)	Piano	.wav
	Violin
	Mixture
MIDI	Piano	.mid
Score	Score Image	.pdf
	Score XML	.musicxml
MIDI‑score alignment	Piano	.match
Note annotations	Violin	.csv
Beat annotations	Mixture	.csv

An example from the KRAISLER dataset with the piece Kreisler’s Liebesfreud, including **(a)** a piano–violin score excerpt, **(b)** audio waveforms with beat positions by red lines (solid: downbeat, dash: beat), and **(c)** a piano roll showing directly captured piano MIDI and estimated violin note annotations.

Pitch histograms of piano and violin note events in the KRAISLER dataset. White and black bars correspond to the piano’s white keys and black keys, respectively. Hatched bars in the lower portion correspond to the violin part.

Note density in notes per second for the piano and violin of each piece in the KRAISLER dataset.

Note activation time of each piece in the KRAISLER dataset.

A comparison of the tempo variability of two pieces in our dataset. **(a)** 2. Kreisler’s Preghiera ( $c_{var} = 0.29$ ) **(b)** 13. Beethoven’s Violin Sonata No. 5 ( $c_{var} = 0.05$ ).

Tempo distribution for each piece with coefficient of variance ( $c_{var}$ ).

Table 4

Automatic piano transcription results evaluated on the piano solo tracks of the KRAISLER and MAESTRO datasets.

Model	MAESTRO			Frame			Note Onset			Note w/ Offset			Note w/ Offset & Vel.
Model	v2	v3	Aug.	P	R	F1	P	R	F1	P	R	F1	P	R	F1
KRAISLER piano solo tracks
OaF		✓		97.7	66.6	78.7	97.3	86.6	91.3	66.8	59.9	63.0	65.1	58.5	61.5
HPPNet		✓		97.6	75.5	84.7	97.5	95.8	96.6	71.2	70.0	70.6	69.1	68.0	68.5
Transkun		✓		96.1	84.5	89.6	95.7	97.3	96.5	77.2	78.4	77.8	74.9	76.0	75.4
Kong	✓			93.1	86.6	89.4	96.8	96.1	96.4	77.2	76.6	76.9	75.7	75.2	75.4
Transkun_Aug		✓	✓	97.8	84.2	90.1	99.7	97.3	98.5	80.1	78.3	79.2	79.2	77.5	78.3
MAESTRO piano solo tracks
Transkun		✓		95.8	95.0	95.4	99.5	97.2	98.3	94.6	92.4	93.5	94.1	91.9	92.9

Table 5

Automatic piano transcription results evaluating on all mixtures of piano and violin duets from our dataset.

Model	MAESTRO			Frame			Note Onset			Note w/ Offset			Note w/ Offset & Vel.
Model	v2	v3	Aug.	P	R	F1	P	R	F1	P	R	F1	P	R	F1
OaF		✓		85.3	44.0	57.3	71.1	74.0	71.6	28.4	29.3	28.5	26.7	27.5	26.8
HPPNet		✓		78.7	57.8	66.2	51.0	88.4	63.8	26.9	46.6	33.6	25.0	43.4	31.3
Transkun		✓		79.6	69.5	73.9	43.6	91.3	58.3	24.1	51.5	32.4	21.3	45.5	28.7
Kong	✓			82.4	85.4	83.5	89.6	91.4	90.3	65.3	66.8	66.0	61.2	62.7	61.9
Transkun_Aug		✓	✓	95.5	78.6	85.6	95.7	94.5	95.0	72.2	71.4	71.7	70.7	69.9	70.2

Table 6

Violin transcription results using the MUSC (Tamer et al., 2023) and VioPTT (Wang et al., 2026) transcription models on solo violin tracks from KRAISLER and representative datasets containing violin recordings.

Testset	MUSC				VioPTT
Testset	P	R	F1	F1_no	P	R	F1	F1_no
URMP	86.5	83.1	84.6	93.0	86.1	83.6	84.5	93.1
Bach10	65.0	64.8	64.8	77.0	68.1	71.8	69.9	79.5
MOSA	59.4	57.6	58.3	72.2	–	–	–	–
KRAISLER	39.4	43.7	40.9	60.5	60.5	53.8	56.7	75.6

Table 7

Comparison of piano source separation results in classical music.

Dataset	Inst.	SDR	SIR	SAR
KRAISLER (Ours)	Piano	9.48 ± 2.64	21.78 ± 4.17	9.87 ± 2.59
	Other (Violin)	9.31 ± 2.60	14.90 ± 4.02	11.09 ± 1.98
PCD (Özer et al., 2023)	Piano	8.52 ± 3.97	12.31 ± 3.66	11.60 ± 4.46
	Other (Orchestra)	4.92 ± 2.90	10.69 ± 3.54	7.02 ± 2.73

Table 8

Beat‑level audio‑to‑score alignment results for piano– violin ensemble mixes across dry, studio, and hall reverb conditions.

Method	AAE (ms) $\pm σ$	MAE (ms)	Alignment Rate (%)
			≤0.3 s	≤0.5 s	≤1.0 s
[Dry mix]
Offline	55.4 ± 108.6	27.2	96.6	98.2	99.7
Online	294.0 ± 378.9	138.9	68.2	77.9	88.8
[Studio mix]
Offline	55.0 ± 105.1	29.5	96.9	98.4	99.6
Online	304.7 ± 381.7	154.3	67.5	77.4	88.9
[Hall mix]
Offline	61.4 ± 117.3	29.9	96.6	98.2	99.5
Online	318.8 ± 376.7	170.5	63.8	75.3	87.8

Table 9

Beat‑tracking results of the KRAISLER dataset.

Model	Accuracy (%)
Model	P	R	F1
TCN (with DBN)	21.2	43.1	27.2
All‑in‑One (with DBN)	67.9	55.7	58.3
RNN+DBN	68.5	70.1	67.1
RNN+PLPDP	53.1	92.6	65.3
RNN+SPPK	70.7	76.1	71.0

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.5334/tismir.338 | Journal eISSN: 2514-3298

Journal RSS Feed

Language: English

Page range: 456 - 473

Submitted on: Sep 1, 2025

Accepted on: May 18, 2026

Published on: Aug 6, 2026

Published by: Ubiquity Press

In partnership with: Paradigm Publishing Services

Keywords:

real performance dataset,

piano transcription,

violin transcription,

piano source separation,

score alignment,

beat tracking

© 2026 Hyemi Kim, Jiyun Park, Sein Lee, Taegyun Kwon, Sunjae Won, Juhan Nam, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.

Volume 9 (2026): Issue 1

KRAISLER: A Multi‑Track Dataset of Piano and Violin Duet Recordings for Music Information Retrieval Research

Figures & Tables

Table 1

Table 2

Figure 1

Table 3

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7