POP909-CL: Human-Corrected Pop Annotations

• POP909-CL is a rigorously human-curated pop music dataset that replaces rule-based annotations with expert-verified labels.
• It corrects misalignments, missing key changes, and tempo noise to provide reliable data for precise chord recognition and MIR analysis.
• The dataset features 909 MIDI files with fixed-tempo scores, enabling improved training and benchmarking of state-of-the-art symbolic music models.

POP909-CL is a rigorously human-curated symbolic pop music dataset providing professionally verified annotations for chords, beats, key-signature events, and time-signature events, derived as a complete quality-controlled refinement of the original POP909 collection. Engineered to address the limitations of algorithmic annotation—namely large-scale misalignments and missing metrical information—POP909-CL enables high-precision symbolic chord recognition and other Music Information Retrieval (MIR) tasks that require reliable symbolic ground-truth.

1. Motivation and Scope

The original POP909 dataset (Liu et al., 2020) comprises 909 MIDI files of Chinese pop songs with rule-based, automatically extracted annotations for beats, chords, keys, and time signatures. While widely adopted, POP909's algorithmic annotations are compromised by systematic misalignments, omission of key or time-signature changes, and tempo noise that undermine its suitability for high-precision chord recognition or detailed symbolic analysis. POP909-CL directly addresses these limitations by substituting all rule-based data with human-validated ground-truth, while preserving the original repertoire and MIDI content.

POP909-CL's scope encompasses:

• 909 piano-arrangement MIDI tracks, each with annotations corrected for chord segments (root, quality, bass), beat onset alignment, explicit key signature events (including mid-piece modulations), time-signature events, and globally flattened (fixed) tempo.
• Systematic correction and validation, ensuring every beat and chord boundary matches musical intent and score, with all pieces now rendered as fixed-tempo scores suitable for time quantization and symbolic model training.

2. Human Annotation and Correction Workflow

Expert musicians and music theorists performed a two-pass manual annotation process:

1. Review method: Annotators interacted with each track within a Digital Audio Workstation (DAW), observing synchronized MIDI piano-roll and score displays. This method allowed precise auditory and visual matching.
2. Correction steps:
   • Beat onsets were re-annotated to ensure strict alignment with metrical structure; ambiguous or misaligned "start beats" from the original dataset were snapped to the correct grid.
   • Chord segments were exhaustively re-segmented, labeling root, quality, and bass with precise boundaries at musical harmonic changes.
   • Key-signature changes, especially mid-piece tonic shifts, were inserted or corrected to reflect the actual tonal structure present in the score.
   • Time-signature switches (e.g., 4/4 to 2/2 or 6/8) were annotated wherever metric or phrasing changes dictated them.
   • All MIDI files were re-tempo-mapped to play back at a single notated tempo, erasing intra-piece tempo fluctuations for fixed-frame symbolic processing.

Quantitative impacts of human correction:

• 40.6% of the original "start beats" were realigned to correct metric positions.
• 14.2% of key-signature changes missing from the initial annotations were added.
• 2.6% of time-signature events, previously wrong or absent, were corrected.
• All such errors are eliminated in POP909-CL, though no absolute error metric beyond these percentages is defined (Yao et al., 8 Oct 2025).

3. Dataset Characteristics

Summary statistics:

• Number of pieces: 909 pop-song MIDI files.
• Aggregate duration: Not explicitly reported; typical median per-track duration is approximately 5 minutes, totaling 10–15 hours of music.
• Chord label structure: Each chord is encoded as (root, quality, bass), where root and bass are 12-class pitch categories, and quality covers 9–12 standard classes (maj, min, 7, m7, sus, dim, aug, etc.), as per the original POP909 schema.
• Time signature: The majority of tracks are in 4/4 (common time), with 2–3% exhibiting mid-piece transitions to 2/2 or 6/8.
• Beat density: Each piece contains approximately 200–300 beats post-correction, calculated at typical tempos and durations.
• Chord segment distribution: Major triads account for ≈55% of all segments, minor triads ≈25%, dominant/seventh variants ≈15%, remainder comprising sus, dim, aug, and added-tone chords (~5%), with detailed histograms furnished in the POP909-CL metadata.

4. Data Representation and File Structure

POP909-CL follows conventions suitable for symbolic MIR research, with the following organizational schema:

• MIDI files: Each song is represented as a standard 16-track MIDI file, playing at a fixed, score-level tempo.
• Annotation files: For every song, the following annotation types are delivered, typically in CSV or JSON formats:
  • Beat list: Onset times or tick indices of each metric beat.
  • Chord list: Ordered tuples $$(t_k, \text{root}_k, \text{quality}_k, \text{bass}_k)$$, with $t_k$ marking the kth chord change event.
  • Key signature list: Tuples $(t_j, ks_j)$ denoting every key signature change, with $ks_j$ specifying the tonic and mode (major/minor).
  • Time signature list: Tuples $(t_l, ts_l)$ for time signature changes (e.g., from 4/4 to 6/8).
• Symbolic score export: Annotations are compatible with MusicXML or midi.xml, embedding explicit direction instruction for tempo, meter, and key events.
• Temporal resolution: Data is quantized to a 12-subdivision-per-beat grid (12 frames per beat), supporting piano-roll tokenization and uniform model input.

5. Comparative Analysis: POP909 vs. POP909-CL

The key improvements in POP909-CL over the original POP909 are summarized in the following table:

Correction Category	POP909 (unrefined)	POP909-CL (human-corrected)
Beat alignment error rate	40.6%	0%
Missing key-signature changes	14.2%	0%
Incorrect/omitted time signatures	2.6%	0%

Chord label consistency: Rule-based chord accuracy (full-chord macro accuracy) on original POP909 is 65.0%. When evaluated using POP909-CL as the reference, state-of-the-art symbolic chord recognition models, including Harmony Transformer v2 and BACHI, achieve 82.3% accuracy, suggesting that approximately 35% of the original POP909 chord labels were inconsistent with expert judgement (Yao et al., 8 Oct 2025). This highlights the extent of annotation noise in the original release and the critical value of the curated labels.

Impact on chord recognition benchmarks: Symbolic ACR models (AugmentedNet, ChordGNN, Harmony Transformer v2, BACHI) demonstrate 75–82% full-chord accuracy when trained and evaluated on POP909-CL, a substantial absolute improvement over results obtained with noisy labels, and a more truthful benchmark of true model capacity.

6. Relevance to MIR Research and Applications

POP909-CL supplies a gold-standard symbolic corpus for MIR research, serving as a high-fidelity pop-music benchmark that complements analogous resources for classical ACR. Its human-verified annotations enable accurate evaluation and training of chord recognition models, facilitate harmonic and metrical analysis, support beat and syncopation studies, and offer a precise substrate for generative modeling under symbolic chord constraints. The clear mitigation of label noise and alignment errors in POP909-CL provides stable, reliable conditions for evaluating fine-grained algorithmic improvements, a necessity in comparative MIR research and for progress in symbolic music learning systems (Yao et al., 8 Oct 2025).