The companion repository does not redistribute the music itself. The 30 s audio excerpts and the beat-reference annotations are sourced from publicly distributed corpora (or, for GlobalMood, public YouTube uploads). This document maps every stem in the paper to its source so anyone can rebuild the audio inputs from scratch.
Stem-level indices:
data/stimuli_index_canonical240.csv — the 240 stems that enter the
validation experiment (Mixed corpora + Ballroom + MIREX +
GlobalMood). The audio_source field is either an
upstream::<corpus> token (look up the row’s corpus below) or a
direct YouTube URL.data/stimuli_index_globalmood200.csv — the 200 GlobalMood stems
in the GlobalTap benchmark. Each stem_id is a
YouTube video ID. audio_source resolves to
https://www.youtube.com/watch?v=<stem_id>. source_country is the
ISO-2 code of the country whose weekly YouTube chart the track was drawn
from (each track charted in one country only): the sampling country, not
the artist’s origin. KR, MX, EG, US, and FR have 20 excerpts each, and 54
other countries share the other 100. The country-panel analyses in
supplementary_analyses/cross_cultural/ use this mapping.For every stem the pipeline operates on a 30 s clip starting at a
corpus-specific offset. That offset is recorded in
config/gt_source_map_canonical240.json (see the clip_offset_s
section of the top-level README) so anyone downloading
full-length audio can extract the matching window.
| Corpus | Audio | Reference annotations |
|---|---|---|
| Ballroom | ISMIR 2004 Ballroom dance corpus (200 clips). | Krebs et al. (2013) annotations redistributed in the Beat This! release. |
| Hainsworth | Hainsworth & Macleod (2003) corpus. | Beat This! release. |
| RWC | RWC Pop / Royalty / Classical / Jazz subsets. | Goto et al. RWC annotations from the bm30_filt clip-window collection. |
| Beatles | Isophonics Beatles corpus. | Mauch et al. Isophonics beats (Beat This! release). |
| Harmonix | Harmonix Set (Nieto et al. 2019). | Harmonix Set beats from the Beat This! release. |
| Tapcorrect | TapCorrect / GTZAN-rhythm pack. | Driedger et al. tap-correct annotations. |
| Candombe | Candombe corpus (Rocamora et al.). | Candombe beat annotations. |
| GTZAN-rhythm | GTZAN with Marchand & Peeters annotations. | Marchand & Peeters beats from the Beat This! release. |
| ASAP | ASAP solo-piano performances. | ASAP score-aligned beats. |
| Carnatic | CompMusic Carnatic music collection. | Srinivasamurthy et al. (2014) beats. |
| Cretan | Holzapfel Cretan-leaping-dance set. | Holzapfel & Stylianou (2011) beats. |
| GuitarSet | GuitarSet (Xi et al. 2018). | Beat This! release. |
| Groove MIDI | Groove MIDI Dataset (Gillick et al. 2019). | Beat This! release. |
| Turkish | Turkish makam recordings (Srinivasamurthy et al. 2014). | Srinivasamurthy et al. (2014) beats. |
| MIREX | MIREX 2006 beat-tracking set (40 clips). | Consensus of the 40 lab annotators’ taps distributed with the MIREX 2006 set, computed with this pipeline (scripts/build_mirex_reference.py). |
source_label = "beat_this_upstream" rows in
config/gt_source_map_canonical240.json mean the reference beats are
those distributed in the
Beat This! release. Everything
else lists its own pack name in the source_label field.
Each stem_id is the YouTube video ID. To rebuild the 30 s clips:
yt-dlp -x --audio-format wav -o "audio/%(id)s.%(ext)s" "<stem-url>"
ffmpeg -ss 30 -t 30 -i audio/<stem-id>.wav -ac 1 -ar 44100 \
audio/globalmood_30s/<stem-id>.wav
(GlobalMood clips start 30 s into each track.)
De-identified tap onsets for the 200-excerpt benchmark are in data/taps/.
Raw session exports with platform identifiers and per-session metadata are not
included (see the paper’s Ethics Statement).
Prolific recruited participants for the eleven benchmark exports and the
listener-rating study. The 40-excerpt GlobalMood validation set also included
three Cint exports, which contributed 58 of its 296 tappers. The batch names
appear in scripts/compute_istc_per_stem.py. Both recruitment sources used the
same PsyNet tapping task and audio-marker checks, and their data were combined
for the analysis.
Figures and per-stem metrics regenerate from data/per_stem_metrics.csv,
data/example_stem_panels.npz, and data/country_panels.csv. Stage 4
re-runs from data/stage_inputs/. To recompute ISTC or Stages 1–3 from raw
taps, point PROJECT_ROOT at a tree that mirrors raw_tapping_data/ and run
scripts/compute_istc_per_stem.py or ./run_pipeline.sh.
The n_participants column is the number of tappers used for each stem’s
reference-free measures, not the study’s overall participant count. Use the
supplementary participant-flow counts when reporting study participation, and
n_participants when interpreting istc or cv_pipeline.
For the 40 Ballroom validation excerpts, an early export of 19 participants’
trials was merged together with the full export, so their trials enter the
pipeline twice. The released annotations, cv_pipeline, and n_participants
come from that input (Ballroom n_participants 39–91, or 39–83 without the
copies), and the released annotations are the ones listeners rated. istc is
computed from the raw merges, which have no copies. Without the copies, 37 of the
40 annotations are effectively unchanged (F = 1.00 against the released beats),
one shifts to the opposite phase, and two change metrical level.
All source corpora retain their original licenses. Outside of
audio_demos/ (eight short case-study clips, see its README.md), this
repository only redistributes derived per-stem annotation outputs
(crowd beats, KDE peaks, F-measure tables) and aggregate per-stem metrics.
No other audio is committed.