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 Key Detection Results

From MIREX Wiki

Submissions[edit]

Submission Title PDF Authors
Baseline: CN1 QM Key Detector (QM Vamp Plugins 1.7.1) Link
ChordyV-Key ChordyV: A Symbolic Harmonic Progression and Tonal Audit System for Audio Key Detection PDF Mulfie Ansyori Sofyan
Chord-LLM Key Detection via LLM Processing of Chord Recognizer Outputs PDF Jeffrey Gordon
Hurliman_K10 K10: A Compact Convolutional Network for Global Key Estimation PDF John Hurliman
MusicFM-MLP Fully Fine-Tuned MusicFM with a Lightweight MLP Head for Global Key Detection PDF Xinwei Jiang, Yaolong Ju
RESON An Unsupervised Spectral Pitch-Class Baseline for Audio Key Detection PDF Ziyi Xu
UFAL-PMCG-music-flamingo-reasoning-full UFAL-PMCG Systems at the MIREX 2026 Audio Key Detection Challenge: Audio-LLMs Are No Longer Off-Key PDF Tomáš Sourada, Jan Hajič jr.
UFAL-PMCG-qwen3-omni-30b-direct-full UFAL-PMCG Systems at the MIREX 2026 Audio Key Detection Challenge: Audio-LLMs Are No Longer Off-Key PDF Tomáš Sourada, Jan Hajič jr.
UFAL-PMCG-qwen25-omni-7b-direct-full UFAL-PMCG Systems at the MIREX 2026 Audio Key Detection Challenge: Audio-LLMs Are No Longer Off-Key PDF Tomáš Sourada, Jan Hajič jr.

Test Sets[edit]

  • GiantSteps Key: 604 two-minute excerpts of electronic dance music (the MIREX key test set since 2017).
  • GTZAN: the 837 GTZAN excerpts with a single key in the annotations of Kraft and Lerch (classical and excerpts marked as modulating or unknown are excluded).
  • FMAKv2 (new in 2026): 5,489 full-length Creative Commons songs of the Free Music Archive in 17 genres, one global key each, annotated by an expert (FMAKv2, Kong et al., ISMIR 2024; audio from FMAK).
  • JSIR (new in 2026): the 226 recordings that stay in one key, out of 267 recordings of jazz standards from the JSIR dataset (Lee and Jeong, 2026). The label is the key in which the standard is played; recordings without a key label, modal tunes, Giant Steps and recordings that change key are left out. Whole recordings are evaluated.
  • Yamaha_JPop: A private dataset annotated by Yamaha Corporation (174 JPOP songs).
  • Yamaha_Balanced: A private dataset annotated by Yamaha Corporation (185 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%).

Results[edit]

Weighted score: 1 for the correct key, 0.5 for a perfect fifth, 0.3 for the relative key, 0.2 for the parallel key, 0 otherwise. Weighted (sym.) gives 0.5 to a fifth above or below the reference (as in MIREX 2017–2019); Weighted (asym.) only to a fifth above (mir_eval 0.8.2 default). The other columns are the percentage of files in each category.

For the two weighted scores and Correct, the best result is shown in bold and the second best underlined. Entries marked * were trained or tuned on that dataset and are not ranked on it.

GiantSteps Key[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM 74.40 72.42 68.05 3.31 3.97 5.96 4.64 3.31 10.76
ChordyV-Key 39.74 35.68 28.81 2.98 8.11 10.10 11.76 1.99 36.26
Hurliman_K10 74.57 72.91 67.88 4.30 3.31 6.95 3.97 3.15 10.43
MusicFM-MLP 77.62 76.37 72.19 4.30 2.48 4.14 3.97 3.97 8.94
RESON 42.30 41.23 29.30 13.08 2.15 5.79 18.21 2.48 28.97
UFAL-PMCG-music-flamingo-reasoning-full 60.17 59.17 52.15 3.15 1.99 10.76 11.09 1.32 19.54
UFAL-PMCG-qwen25-omni-7b-direct-full 56.54 53.23 45.86 8.94 6.62 7.78 2.81 1.99 25.99
UFAL-PMCG-qwen3-omni-30b-direct-full 70.94 68.79 62.91 6.13 4.30 6.29 4.64 2.48 13.24
Baseline: CN1 50.53 46.97 39.74 4.80 7.12 13.24 4.30 2.65 28.15

GTZAN[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM 73.76 72.09 67.03 3.70 3.35 2.51 12.31 0.84 10.27
ChordyV-Key 48.64 43.44 36.44 5.73 10.39 5.26 12.78 1.67 27.72
Hurliman_K10 72.03 70.24 64.64 4.90 3.58 5.02 8.24 1.79 11.83
MusicFM-MLP 73.42 71.80 66.91 2.15 3.23 3.82 13.38 1.19 9.32
RESON 48.48 47.35 34.29 16.97 2.27 5.62 14.46 2.63 23.78
UFAL-PMCG-music-flamingo-reasoning-full 54.15 52.77 44.68 4.90 2.75 7.17 17.44 0.36 22.70
UFAL-PMCG-qwen25-omni-7b-direct-full 55.94 52.47 45.04 7.53 6.93 7.53 7.05 1.55 24.37
UFAL-PMCG-qwen3-omni-30b-direct-full 67.60 64.43 58.90 4.54 6.33 5.62 7.89 2.87 13.86
Baseline: CN1 47.08 43.14 36.68 4.18 7.89 8.84 8.60 6.69 27.12

FMAKv2[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM* 77.60 75.73 71.51 3.64 3.73 5.68 3.48 2.57 9.38
ChordyV-Key 53.40 47.11 41.12 5.21 12.57 8.11 4.77 1.95 26.27
Hurliman_K10 72.43 70.33 64.40 5.56 4.21 7.62 4.32 1.55 12.35
MusicFM-MLP 75.63 73.79 68.46 4.39 3.68 8.31 3.21 2.17 9.78
RESON 56.66 55.45 43.58 15.49 2.42 8.09 8.53 1.13 20.77
UFAL-PMCG-music-flamingo-reasoning-full 69.52 67.41 61.14 5.72 4.23 8.58 4.15 1.82 14.36
UFAL-PMCG-qwen25-omni-7b-direct-full 66.00 63.25 56.33 6.74 5.50 9.75 3.12 2.00 16.56
UFAL-PMCG-qwen3-omni-30b-direct-full 74.86 72.69 67.41 4.28 4.35 7.76 4.04 1.88 10.28
Baseline: CN1 63.02 58.91 52.94 4.01 8.22 10.49 4.10 3.15 17.09

* Chord-LLM: according to its extended abstract, the system was developed and tuned on FMAKv2. Its FMAKv2 outputs were produced by the author to save API costs (not run by the organizers) and scored by the organizers.

JSIR[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM 76.90 75.58 68.58 6.64 2.65 8.41 5.75 0.44 7.52
ChordyV-Key 53.81 46.28 39.82 5.31 15.04 9.73 4.42 0.44 25.22
Hurliman_K10 34.78 33.67 21.68 3.98 2.21 25.66 11.50 0.00 34.96
MusicFM-MLP 79.82 76.50 71.68 2.21 6.64 11.50 1.33 0.44 6.19
RESON 31.81 31.81 18.14 19.03 0.00 12.39 2.21 0.89 47.35
UFAL-PMCG-music-flamingo-reasoning-full 70.04 68.94 59.29 10.18 2.21 11.06 6.19 0.00 11.06
UFAL-PMCG-qwen25-omni-7b-direct-full 62.74 58.54 51.33 3.54 8.41 9.29 13.27 0.00 14.16
UFAL-PMCG-qwen3-omni-30b-direct-full 63.54 62.65 53.54 2.21 1.77 17.26 14.16 0.00 11.06
Baseline: CN1 53.05 49.29 42.92 2.21 7.52 11.95 8.41 6.64 20.35

YAMAHA_Balanced[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM 91.46 90.38 88.11 0.54 2.16 5.95 1.08 0.54 1.62
ChordyV-Key 76.16 70.22 64.86 4.86 11.89 8.65 1.62 0.00 8.11
Hurliman_K10 89.30 89.30 85.95 0.54 0.00 8.11 3.24 0.00 2.16
MusicFM-MLP 91.68 91.68 89.19 0.00 0.00 6.49 2.70 0.00 1.62
RESON 62.70 62.43 47.03 24.32 0.54 6.49 6.49 0.54 14.59
UFAL-PMCG-music-flamingo-reasoning-full 90.59 90.32 87.03 1.62 0.54 7.57 1.08 0.00 2.16
UFAL-PMCG-qwen25-omni-7b-direct-full 79.51 78.16 72.97 4.32 2.70 7.57 3.78 0.00 8.65
UFAL-PMCG-qwen3-omni-30b-direct-full 90.81 90.81 88.11 0.54 0.00 7.03 1.62 0.00 2.70
Baseline: CN1 83.03 81.41 77.84 0.00 3.24 10.81 1.62 2.70 3.78

YAMAHA_JPop[edit]

Group Weighted (sym.) Weighted (asym.) Correct Fifth up Fifth down Relative Parallel Chromatic Other
Chord-LLM 91.49 90.92 87.36 1.15 1.15 9.20 1.15 0.00 0.00
ChordyV-Key 63.62 59.89 52.87 2.30 7.47 18.39 1.72 0.00 17.24
Hurliman_K10 85.80 85.80 79.89 1.15 0.00 16.67 1.72 0.00 0.57
MusicFM-MLP 87.99 87.99 83.33 0.57 0.00 13.79 1.15 0.00 1.15
RESON 55.69 55.11 37.93 22.41 1.15 17.24 4.02 0.00 17.24
UFAL-PMCG-music-flamingo-reasoning-full 85.86 85.57 79.89 1.72 0.57 16.09 0.00 0.00 1.72
UFAL-PMCG-qwen25-omni-7b-direct-full 77.53 76.95 70.69 2.30 1.15 16.67 0.57 1.72 6.90
UFAL-PMCG-qwen3-omni-30b-direct-full 87.47 87.47 82.18 1.15 0.00 14.94 1.15 0.00 0.57
Baseline: CN1 75.63 75.06 67.82 1.15 1.15 21.84 0.57 0.57 6.90

Notes[edit]

  • MusicFM-MLP was trained on the Hooktheory annotations released with SheetSage, which may overlap with some popular song datasets to a small extent.