Read-only mirror of the MIREX Wiki, snapshot of 29 October 2026. Open this page on the original wiki to edit it or to see newer changes.

2026:Audio Downbeat Estimation Results

From MIREX Wiki

Submissions[edit]

Submission Title PDF Authors
Baseline: CD4 QM Bar & Beat Tracker (beats per bar = 4) Link
Baseline: BeatThis Beat This! Accurate Beat Tracking Without DBN Postprocessing Link
CASM Context-Aware Semi-Markov Post-Processor for Beat and Downbeat Estimation PDF Zhanhong He, Hanyu Meng, Jiahao Zhao, Ganghui Ru, Yaolong Ju
infgon Multi-Scale Expert Reverification for Joint Beat and Downbeat Tracking PDF Tianle Yang

Test Sets[edit]

  • GTZAN: 993 excerpts of the GTZAN dataset with bar annotations (gtzan_tempo_beat, completed with GTZAN-Rhythm; 6 jazz excerpts without bar annotations are excluded).
  • osu2beat2025 (new in 2026): 703 full-length songs from ranked osu! beatmaps (anime, Vocaloid, game music and more) with metered beat annotations (Osu2MIR, ISMIR 2025 LBD); audio taken from public osu! mirrors and verified against the MD5 given with each annotation.
  • Yamaha_JPop: A private dataset annotated by Yamaha Corporation (250 JPOP songs).
  • Yamaha_Balanced: A private dataset annotated by Yamaha Corporation (239 songs). While it is still biased towards JPOP songs, the dataset covers a wider range of genres: J.Pop (10.37%), Rock (10.37%), J.Enka (10.37%), J.Kayoukyoku (10.37%), Soundtrack (10.37%), Western Pop (10.37%), Children's Song (10.37%), R&B (6.22%), Hiphop (4.56%), Jazz (2.49%), Dance (2.49%), World (2.07%), Techno (1.24%), Easy listening (1.24%), J.Minyou (1.24%), Others (5.81%).
  • JSIR (new in 2026): 260 recordings of jazz standards from the JSIR dataset (Lee and Jeong, 2026), with downbeats annotated by hand. See the notes below the JSIR table.

Metrics: F-measure, precision and recall of the estimated downbeats with a ±70 ms tolerance window, computed per file and averaged over each dataset, as in MIREX 2014–2016 (the first 5 s are not discarded).

For each metric, the best result is shown in bold.

GTZAN[edit]

Group F-measure Precision Recall
CASM beatthis-casm (default) 77.24 77.43 79.51
CASM beatthis-direct 77.33 75.28 82.39
CASM beatthis-dbn55_215 76.56 78.35 76.66
CASM beatthis-dbn30_300 76.88 78.70 77.08
CASM mscnn-casm 70.95 69.42 75.27
CASM mscnn-direct 68.67 63.43 77.91
CASM mscnn-dbn55_215 71.56 73.42 71.46
CASM mscnn-dbn30_300 70.89 73.18 70.67
CASM tcn-casm 65.38 61.91 76.31
CASM tcn-direct 60.75 49.62 85.32
CASM tcn-dbn55_215 66.41 67.51 67.89
CASM tcn-dbn30_300 65.81 67.05 67.18
infgon 13.95 14.19 14.31
Baseline: CD4 46.36 46.43 47.95
Baseline: BeatThis 78.12 75.67 83.93

osu2beat2025[edit]

Group F-measure Precision Recall
CASM beatthis-casm (default) 81.99 87.64 78.48
CASM beatthis-direct 83.78 86.91 82.33
CASM beatthis-dbn55_215 82.19 89.97 77.56
CASM beatthis-dbn30_300 82.38 89.97 77.82
CASM mscnn-casm 71.86 73.57 71.81
CASM mscnn-direct 72.46 71.43 75.07
CASM mscnn-dbn55_215 71.01 81.18 65.74
CASM mscnn-dbn30_300 70.54 81.57 65.16
CASM tcn-casm 63.92 56.82 80.80
CASM tcn-direct 60.89 48.89 85.16
CASM tcn-dbn55_215 67.08 75.88 62.71
CASM tcn-dbn30_300 66.48 75.44 62.23
infgon 11.15 14.18 9.45
Baseline: CD4 44.01 48.77 41.75
Baseline: BeatThis 83.22 86.69 81.77

YAMAHA_Balanced[edit]

Group F-measure Precision Recall
CASM beatthis-casm (default) 85.57 85.21 86.99
CASM beatthis-direct 85.29 82.27 89.91
CASM beatthis-dbn55_215 87.03 86.78 88.12
CASM beatthis-dbn30_300 87.10 86.92 88.11
CASM mscnn-casm 81.10 79.96 83.51
CASM mscnn-direct 78.25 72.74 85.98
CASM mscnn-dbn55_215 83.60 84.03 84.25
CASM mscnn-dbn30_300 83.52 84.42 83.79
CASM tcn-casm 72.69 67.61 85.30
CASM tcn-direct 63.61 50.93 89.50
CASM tcn-dbn55_215 76.29 74.24 81.03
CASM tcn-dbn30_300 75.68 73.59 80.47
infgon 11.40 11.47 11.85
Baseline: CD4 60.45 57.46 66.03
Baseline: BeatThis 86.09 83.39 90.33

YAMAHA_JPop[edit]

Group F-measure Precision Recall
CASM beatthis-casm (default) 91.78 92.22 92.09
CASM beatthis-direct 91.54 90.04 94.07
CASM beatthis-dbn55_215 92.66 93.23 92.83
CASM beatthis-dbn30_300 92.59 93.13 92.83
CASM mscnn-casm 85.64 84.87 87.65
CASM mscnn-direct 83.49 78.60 90.36
CASM mscnn-dbn55_215 88.40 89.68 88.43
CASM mscnn-dbn30_300 88.46 89.98 88.36
CASM tcn-casm 79.48 75.41 89.09
CASM tcn-direct 70.11 57.20 94.48
CASM tcn-dbn55_215 83.27 82.50 86.52
CASM tcn-dbn30_300 82.77 81.84 86.51
infgon 12.36 12.60 12.59
Baseline: CD4 63.10 60.33 68.86
Baseline: BeatThis 91.46 90.33 93.67

JSIR[edit]

Group F-measure Precision Recall
CASM beatthis-casm (default) 85.52 88.76 83.67
CASM beatthis-direct 84.27 81.14 88.52
CASM beatthis-dbn55_215 77.67 89.79 71.69
CASM beatthis-dbn30_300 83.02 90.94 78.64
CASM mscnn-casm 75.94 72.56 81.69
CASM mscnn-direct 69.36 61.17 81.73
CASM mscnn-dbn55_215 70.16 83.54 64.52
CASM mscnn-dbn30_300 71.27 83.72 65.91
CASM tcn-casm 54.47 44.31 82.09
CASM tcn-direct 48.96 35.48 83.31
CASM tcn-dbn55_215 54.52 60.31 54.79
CASM tcn-dbn30_300 54.16 59.74 54.48
infgon 10.60 13.97 9.32
Baseline: CD4 27.82 31.12 27.05
Baseline: BeatThis 87.27 84.17 91.53

JSIR is an experimental beat and downbeat test set this year. Its labels are downbeats marked by hand on beats that were extracted automatically with madmom and then corrected by hand; none of the systems above uses madmom's pretrained models (the CASM dbn variants only use madmom's DBN decoder on their own network outputs). Besides, temporary settings for this year:

  • 7 recordings that overlap the training data of Beat This! and CASM (via JAAH, Hainsworth and SIMAC) are removed.
  • Metrics are evaluated within the annotated (valid) segments only.