Submissions[edit]

Interactive Demo[edit]

The chords of every 2026 system on two Creative Commons songs released in October 2026, compared with a reference annotation at each evaluation level, are available on the interactive demo page. Click a chord to hear it on a piano, or let the piano play a system's chords along with the song.

Test Sets[edit]

Main Test Sets[edit]

The following datasets are served as pure test sets. No system is allowed to train on them.

• Billboard 2013: The held-out portion of the McGill Billboard dataset (161 songs), containing mainly western pop songs from the Billboard chart.
• JSIR (new in 2026): 267 recordings of 44 jazz standards from the JSIR dataset (Lee and Jeong, 2026), with chord annotations derived from the lead sheets of the standards and aligned to the recordings. Only the valid segments marked by the dataset authors are evaluated (rubato introductions, unaccompanied or ambiguous passages are left out).
• Yamaha_JPop: A private dataset annotated by Yamaha Corporation (249 JPOP songs).
• Yamaha_Balanced: A private dataset annotated by Yamaha Corporation (228 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%).

Additional Test Sets[edit]

These are datasets that may not be strictly held-out test sets. Some models might have been trained on these datasets; for specific details, please refer to the extended abstracts of each model.

• Billboard 2012: The public portion of the McGill Billboard dataset (188 songs).
• RWC Popular: 100 pop songs from the RWC (Real World Computing) Music Database. 20% songs with English lyrics and 80% songs with Japanese lyrics.
• JAAH (new in 2026): 113 jazz recordings of the Jazz Audio-Aligned Harmony dataset, with the official published chord annotations.

Main Results[edit]

The following datasets are served as pure test sets. No system is allowed to train on them.

For each metric, the best result is shown in bold. Entries marked * were trained or tuned on that dataset and are not ranked on it.

Billboard2013[edit]

JSIR[edit]

SLASH: decoder settings were tuned on development data that included JAAH songs; 5 of the 267 JSIR recordings are also in JAAH.

YAMAHA_Balanced[edit]

YAMAHA_JPop[edit]

Additional Results[edit]

Below are results on datasets that may not be strictly held-out test sets. Some models might have been trained on these datasets; for specific details, please refer to the extended abstracts of each model.

Billboard2012[edit]

* Trained on this dataset: SLASH (public McGill Billboard annotations); Baseline ChordFormer (the 1,217-song set of Humphrey and Bello, which includes the public McGill Billboard songs).

RWC-Popular[edit]

* Trained on this dataset: SLASH (MARL-Chords annotations of the RWC Popular songs); MuQ-BTC (RWC-Pop, 80% of the songs in the submitted fold); Baseline ChordFormer (the 1,217-song set of Humphrey and Bello, which includes RWC Popular).

JAAH[edit]

* SLASH: decoder settings were tuned on development data that included JAAH songs (the network itself was not trained on JAAH).

Notes[edit]

• MuQ-BTC was trained on SheetSage-Chord, which may overlap with some popular song datasets to a small extent.

Update (2026-10-16): MuQ-BTC scores on Billboard2013 and JAAH changed (by up to 0.03) after a fix in the post-processing of the X label; JSIR results added.