STAR Drums: A Dataset for Automatic Drum Transcription

Philipp Weber; Christian Uhle; Meinard Müller; Matthias Lang

doi:10.5334/tismir.244

Figures & Tables

Overview of the STAR Drums creation process. The non‑drum stem and the original drum stem are obtained from the original mix by using an MSS algorithm. An ADT algorithm creates an estimated annotation from the original drum stem, which is then used to render the re‑synthesized drum stem by using virtual drum instruments. The re‑synthesized drum stem is mixed with the non‑drum stem, resulting in the STAR mix, which forms the STAR Drums dataset together with the estimated annotation, which is now regarded as the reference annotation.

Table 1

Overview of available ADT datasets.

Dataset	Non‑drum Instr.	Drums	Vocals	Melodic Instr.	# Drum Classes	Len. [h]
RWC Music Database (Goto et al., 2002)	Rec.	Rec.	Yes	Yes	29	18.1
ENST Drums (Gillet and Richard, 2006)	Rec.	Rec. & Synth.	No	Yes	20	1.0
MDB Drums (Southall et al., 2017)	Rec.	Rec.	Yes	Yes	20	0.4
RBMA13 (Vogl et al., 2017)	Rec.	Rec.	Yes	Yes	23	1.9
TMIDT (Vogl et al., 2018)	Synth.	Synth.	No	Yes	18	257.1
Slakh (Manilow et al., 2019)	Synth.	Synth.	No	Yes	—	118.3
A2MD (Wei et al., 2021)	Rec.	Rec.	Yes	Yes	3	34.5
ADTOF‑RGW (Zehren et al., 2021)	Rec.	Rec.	Yes	Yes	5	89.2
ADTOF‑YT (Zehren et al., 2023)	Rec.	Rec.	Yes	Yes	5	202.2
Proposed STAR Drums	Rec.	Synth.	Yes	Yes	18	124.5

[i] In Slakh, no mapping from MIDI notes to drum classes is provided. Therefore, the number of supported classes depends on the mapping created by the user.

MIDI note velocity distribution of all drum classes and all tracks.

Table 2

Input data for STAR Drums.

Dataset	Instrument stems provided	# Total Tracks	# Used Tracks	Len. Used Tracks [h]
MUSDB18 (Rafii et al., 2017)	Yes	150	150	9.8
ISMIR04 (Cano et al., 2006)	No	2000	1228	98.4
MTG‑Jamendo (Bogdanov et al., 2019)	No	55000	4807	302.9

Table 3

Splits of STAR Drums.

Split	Origin of data	MSS algorithm applied	Full tracks	Len. [h]
Training	ISMIR04	Yes	No	20.6
Training	MTG‑Jamendo	Yes	No	94.1
Training (total)	ISMIR04 + MTG‑Jamendo	Yes	No	114.7
Validation	MUSDB18	No	Yes	8.3 (6.7)
Test	MUSDB18	No	Yes	1.6 (0.3)

[i] Values in brackets indicate the duration of audio files that users must create by executing a mixing script. This is necessary because some track licenses of MUSDB18 do not permit the redistribution of remixed versions.

Table 4

Drum classes used with mapping to eight‑, five‑, and three‑class vocabulary, based on Vogl et al. (2018) and Zehren et al. (2023).

Class name	# Classes
	18	8	5	3
Bass drum	BD	BD	BD	BD
Snare drum	SD	SD	SD	SD
Side stick	SS
Hand clap	CLP
Closed hi‑hat	CHH	HH	HH	HH
Pedal hi‑hat	PHH
Open hi‑hat	OHH
Tambourine	TB
Low tom	LT	TT	TT
Mid tom	MT
High tom	HT
Splash cymbal	SPC	CY	CY
Chinese cymbal	CHC
Crash cymbal	CRC
Ride cymbal	RD	RD
Ride bell	RB	BE
Cowbell	CB	BE
Clave/sticks	CL	CL

Genre distribution of the STAR Drums dataset.

Relative class frequencies of STAR Drums, MDB Drums, ENST Drums, and RBMA13.

Total number of detected drum sounds when transcribing ideal and non‑ideal non‑drum stems and ideal and non‑ideal drum stems of MUSDB18.

Table 5

Global F‑measure when training with TMIDT, Slakh, ADTOF, or STAR Drums and testing with MDB Drums, ENST Drums, RBMA13, or STAR Drums for models transcribing 3, 5, 8, and 18 classes.

Model		Test Datasets
CL		MDB Drums	ENST Drums	RBMA13	STAR Drums Test
3	TMIDT	0.78	0.71	0.62	0.72
	Slakh	0.76	0.77	0.55	0.73
	STAR Drums	0.81	0.78	0.67	0.85
	ADTOF‑RGW	0.80	0.80	0.67	0.77
	ADTOF‑YT	0.83	0.79	0.62	0.72
5	TMIDT	0.65	0.69	0.55	0.61
	Slakh	0.68	0.72	0.48	0.59
	STAR Drums	0.79	0.77	0.62	0.82
	ADTOF‑RGW	0.78	0.75	0.60	0.72
	ADTOF‑YT	0.79	0.76	0.59	0.66
8	TMIDT	0.63	0.66	0.52	0.63
	Slakh	0.66	0.71	0.47	0.61
	STAR Drums	0.75	0.74	0.61	0.80
18	TMIDT	0.58	0.61	0.41	0.55
	Slakh	0.59	0.63	0.39	0.58
	STAR Drums	0.67	0.66	0.50	0.78

Global F‑measure and F‑measure per instrument on MDB Drums for five classes when training with TMIDT, Slakh, ADTOF‑RGW, ADTOF‑YT, STAR Drums, and the original mix of STAR Drums (see Section 4.2.4). Class abbreviations are explained in Table 4.

Global F‑measure and F‑measure per instrument on MDB Drums for 18‑class vocabulary when training with TMIDT, Slakh, STAR Drums, and the original mix of STAR Drums (see Section 4.2.4). The classes hand clap, cowbell, and clave/sticks are excluded as MUSDB18 does not contain annotations for these classes. Class abbreviations are explained in Table 4.

Table 6

Global F‑measure results for training with the original mix and estimated annotations (pseudo‑labels), STAR random mix that combines re‑synthesized drum stems and non‑drum stems from different tracks, and a combination of both methods when transcribing 5 and 18 classes.

Model		Test Datasets
CL	STAR DrumsTraining Data	MDB Drums	ENST Drums	RBMA13	STAR Drums Test
5	STAR mix	0.79	0.77	0.62	0.82
	Original mix	0.78	0.80	0.60	0.73
	STAR random mix	0.78	0.77	0.61	0.81
	STAR mix + original mix	0.78	0.78	0.61	0.77
	STAR mix + STAR random mix	0.77	0.74	0.60	0.80
	Original mix + STAR random mix	0.78	0.78	0.60	0.77
18	STAR mix	0.67	0.66	0.51	0.78
	Original mix	0.66	0.72	0.53	0.65
	STAR random mix	0.65	0.67	0.51	0.77
	STAR mix + original mix	0.68	0.70	0.54	0.73
	STAR mix + STAR random mix	0.67	0.66	0.51	0.78
	Original mix + STAR random mix	0.67	0.69	0.53	0.73

[i] For comparison, results using the STAR mix are also provided.

STAR Drums: A Dataset for Automatic Drum Transcription

Figures & Tables

Figure 1

Table 1

Figure 2

Table 2

Table 3

Table 4

Figure 3

Figure 4

Figure 5

Table 5

Figure 6

Figure 7

Table 6

Paradigm

My account