Itsuki-music/POP909-CL-Dataset

POP909-CL Dataset

Paper | Demo Page | POP909-CL Dataset | Model Weight & Data

POP909-CL is a chord-labelled extension of the original POP909 corpus, intended to support chord recognition, automatic accompaniment, and music theory research. The repository combines expert-reviewed chord annotations with metadata-corrected MIDI files while preserving backwards compatibility with the POP909 24-bin timing grid. It is part of the release of the paper "BACHI: Boundary-Aware Symbolic Chord Recognition Through Masked Iterative Decoding on Pop and Classical Music". You can refer to details about this project here

TL;DR

For anyone want to just use the data, please go to POP909_processed/ subfolder (or POP909_processed.zip), which contains corrected MIDI with chord labels, and POP909_processed/metadata.csv for per-piece metadata.

Dataset Highlight

Key Improvements

• ✅ 40.6% of misaligned start beats corrected
• ✅ 14.2% of missing key signature changes added
• ✅ 2.6% of incorrect time signatures fixed
• ✅ ~35% of chord label errors corrected

Example Correction

Dataset Structure

• POP909_processed/ – The release: one NNN.mid per piece plus metadata.csv / metadata.json.
• POP909_processed.zip – The same folder, zipped.
• POP909_chord_labels/ – One NNN.csv per piece with the chord labels derived from the CHORDS track (see Chord labels).
• POP909_chord_annotated/ – Raw expert-annotated MIDI files (merged score, algorithmic chord track, human chord track). The release takes the human chord track from here.
• annotations/corrections.csv – Per-piece downbeat and meter corrections made after the expert annotation.
• scripts/ – Utilities for checking the release and writing chord markers.
• process_pop909.py – Exports chord labels (chord_symbol.csv) and score-only MIDI for training.

File Format

Each POP909_processed/NNN.mid is a type-1 MIDI file with five tracks:

Track	Content
0 POP909-CL NNN	Tempo, time signature and key signature events, and a marker with the chord label (e.g. C#:m7/E) at each chord onset
1 MELODY, 2 BRIDGE, 3 PIANO	The three POP909 tracks, from the POP909 24-bin quantization, with the original sustain/sostenuto/soft pedals (CC64/66/67)
4 CHORDS	Human-annotated chords, one sustained note per chord tone

• Timing. One POP909 beat is 480 ticks and lasts 0.6 s in every file (the tempo of the 24-bin MIDI).
• Downbeats. Bar lines follow the corrected downbeat. Each file is then moved by whole bars so that the earliest note falls in the first bar. For example, a 4/4 piece starting on beat 7 starts on beat 3.
• PPQ. 892 files use PPQ 480, so a POP909 beat is a quarter note. In the 14 x/8 pieces a POP909 beat is an eighth note, so they use PPQ 960, except 092, whose beat is a dotted quarter (PPQ 320). Tempo is scaled to match, so playback speed and tick positions are the same as everywhere else, and ticks / ticks_per_beat always counts true quarter notes.
• Pedals. Pedal events from the original POP909 MIDI are preserved
• Keys and meters. Read the key and time signature events on track 0, including changes during the piece.

Metadata

metadata.csv has one row per piece (metadata.json adds full detail):

• status: released or missing (see missing pieces).
• initial_key, initial_meter, ppq, key_changes, meter_changes (as key@tick).
• total_shift_ticks: the tick offset from the POP909 24-bin timeline.
• flags: caveats worth checking before relying on a piece (see below).

Missing Pieces

Piece	Reason
553	The annotated score was a copy of 543 (#1); no annotation of the genuine 553 exists
367	The annotated file has no chord track
658	The annotated file has no score track, so its chords cannot be aligned

Known Issues

10 released pieces carry flags in the metadata, e.g. a single bar of a different meter, or an approximate meter change.

What's Revised

• Score. Rebuilt from the POP909 24-bin quantization, keeping the original MELODY / BRIDGE / PIANO tracks and pedals.
• Further corrections. Some downbeat and meter errors the annotation missed were found by automatic detection and spot-checked by listening.
• Issues #1 and #2. 553 (a duplicate of 543) is removed, and 043 .mid is renamed.

Automatic error detection. For every beat we combine simple evidence: POP909's own downbeat estimates, accents in the score (bass, long notes, velocity), harmonic change, and where the annotated chords change. A small model learned from expert-checked pieces then estimates the most likely downbeat position and meter, and flags pieces, or sections of pieces, that disagree with the current bar lines.

Chord Labels

The CHORDS track spells each chord as notes. Labels are derived by simple template matching of the sounding pitch classes against triad and seventh-chord templates; the lowest note gives the bass. POP909_chord_labels/NNN.csv lists offset_qb (onset in quarter notes), root, quality, bass and local_key, and the same labels appear as markers in each MIDI file. To regenerate them:

python process_pop909.py --pop909-root POP909_processed --out data_root/all_data_collection

Citation

If you use this resource, please cite this curated release:

@inproceedings{bachi2025,
    author = {Mingyang Yao, Ke Chen, Shlomo Dubnov, Taylor Berg-Kirkpatrick},
    title = {BACHI: Boundary-Aware Symbolic Chord Recognition Through Masked Iterative Decoding on Pop and Classical Music},
    booktitle = {Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
    year = {2026}
}

along with original POP909 paper:

@inproceedings{pop909-ismir2020,
    author = {Ziyu Wang* and Ke Chen* and Junyan Jiang and Yiyi Zhang and Maoran Xu and Shuqi Dai and Guxian Bin and Gus Xia},
    title = {POP909: A Pop-song Dataset for Music Arrangement Generation},
    booktitle = {Proceedings of 21st International Conference on Music Information Retrieval, {ISMIR}},
    year = {2020}
}