GlobalTap: Globally Diverse Beat Tracking Benchmark via Crowdsourced Tapping
Authors: William T. Botticelli-Wells, Manuel Anglada-Tort, Elif Celen, Rena Far, Harin Lee, Peter M. C. Harrison, Nori Jacoby
Organizations: Cornell University, USA · Goldsmiths, University of London, UK · Max Planck Institute for Empirical Aesthetics, Germany · University of Toronto, Canada · University of Cambridge, UK
Existing beat-tracking benchmarks draw from a narrow range of musical cultures, predominantly Western, and are difficult to scale due to reliance on expert annotation. We address both limits with GlobalTap, a benchmark of 200 popular-music excerpts spanning 59 countries, annotated by 746 participants recruited across ten countries. The benchmark is created using a scalable online tapping pipeline that records taps via audio for high temporal precision, aggregates pooled tap times into consensus beat annotations, and computes intrinsic quality measures that flag perceptually difficult excerpts without requiring a reference. In a separate study on a different 240-excerpt validation set, tapped by an additional 733 participants, an independent group of 130 listeners rated how well the resulting beat tracks lined up with the music, and we detected no overall difference between our annotations and existing reference annotations. Crowd-derived annotations occasionally diverged from the references in phase and metrical interpretation, yet listeners did not consistently prefer the reference versions, suggesting that crowd-derived and existing reference annotations can reflect different perceptual or structural interpretations. The disagreements were often systematic rather than random, and metrical-level choices differed across the ten recruitment countries. We release GlobalTap together with open-source pipeline code as a listener-derived resource for evaluating beat-tracking systems on culturally diverse music.
Figures & tables
Corpus
n
Crowd
Ref
Δ
p
dz
Overall (paired)
200
4.67
4.68
−0.01
.231
−0.02
Ballroom
80
5.01
5.14
−0.13
.007
−0.22
Mixed corpora
100
4.47
4.37
+0.09
.546
+0.10
MIREX
20
4.33
4.41
−0.08
.557
−0.09
GlobalMood
40
4.93
—
—
—
—
Table 1: Per-corpus mean listener ratings (1–7 scale) for the pipeline (Crowd) vs. Reference annotations on the 200 paired validation excerpts plus the 40-excerpt GlobalMood absolute-quality set. Δ=Crowd−Ref ; effect sizes are paired-sample Cohen’s dz ; p -values from Wilcoxon signed-rank on excerpt-level means.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Source corpus
n
Crowd
Ref
Δ
p
pBH
dz
ASAP
8
3.35
3.54
−0.19
.742
.877
−0.20
Beatles
8
4.89
4.07
+0.82
.055
.178
+0.83
Candombe
13
4.40
4.17
+0.23
.733
.877
+0.20
Carnatic
4
4.69
4.84
−0.15
.625
.877
−0.30
Cretan
3
5.36
5.63
−0.27
.500
.812
−0.80
Groove (MIDI)
8
4.45
4.79
−0.33
.312
.580
−0.44
Appendix
Table A1 : Crowd and Reference ratings by source corpus in the mixed corpora. Crowd and Ref are mean per-excerpt ratings. Δ=Crowd−Ref . pBH is corrected across the 13 source-corpus tests.
Excerpt set
Reference source
Excerpts
Tappers
Median per excerpt
Trials
Ballroom
Expert beat grids
80
231
40
3,821
Mixed corpora
13 source corpora
100
104
36
3,555
MIREX
Lab-tapper consensus (40 participants)
20
102
67
1,353
GlobalMood validation
None (absolute-quality check)
40
296
110
4,403
GlobalTap benchmark
Crowd-derived
200
746
117
23,493
Total
—
440
1,479
—
36,625
Appendix
Table A2 : Tapper groups in the benchmark and validation analyses.
Variable
Distribution
Age (years)
M=33.1 , SD=11.1
Gender
49.5% female, 49.1% male, 1.4% non-binary or undisclosed
Formal musical training
Median 1 year (IQR 0–4), with 33% none and 9% at least 10 years
Training median by panel
1–6 years
Gold-MSI (25-item, 1–7)
Overall M=4.10 , SD=0.88 , with panel means 3.77–4.56
Appendix
Table A3 : Demographics of the 746 GlobalTap benchmark tappers.
Period relation
n
Crowd higher
Median Δ
Matched rate
145
51
−.143
Crowd half rate
9
9
+.748
Crowd double rate
14
7
−.238
Other ratio
12
7
+.393
Appendix
Table A4 : Rating differences by Crowd-to-Reference period relation.
Metric
madmom
Beat This!
CMLc
.797
.621
CMLt
.807
.636
AMLc
.947
.908
AMLt
.956
.935
Information gain
.669
.643
Appendix
Table A5 : Beat-tracking agreement with the GlobalTap consensus.
Step
Setting
Search
Period 0.28 – 1.25 s ( 48 – 214 BPM) in 120 steps, phase in 240 steps per period. C is the proportion of peaks within 0.12T+20 ms of a grid beat. The density has an 80 ms Gaussian at each peak, scaled to a maximum of 1
Refinement
Least squares on peaks within 0.15T of the grid, then the phase re-searched within ±0.1T , keeping only phases in [0,T)
Snapping
A beat with a peak within 0.12T moves 45% of the way toward it
Rate doubling
When more than 55% of the midpoints between beats have a peak within 0.18T , the mean density at the midpoints is over 40% of that at the beats, and the doubled rate is at most 214 BPM. The beats of the doubled grid are then nudged toward even spacing
Gap filling
Beats inserted in gaps over 1.4T and out to the window edges, at the median beat interval
Split peaks
Tried when the largest gap between the sorted inter-peak intervals shorter than 0.9T exceeds 0.10T . The grids are compared by 2C′+(1−CV)+0.5Dˉ , where C′ is the proportion of peaks within 0.12T+30 ms of a beat and CV is the coefficient of variation of the beat intervals
Schulich School of Music, McGill University, Montréal, QC, Canada · Centre for Interdisciplinary Research in Music Media and Technology (CIRMMT), Montréal, Canada · Courant Institute, New York University, New York, USA +1