Source-linked AI summary
BACHI: Boundary-Aware Symbolic Chord Recognition Through Masked Iterative Decoding on Pop and Classical Music
Mingyang Yao, Ke Chen, Shlomo Dubnov, Taylor Berg-Kirkpatrick
TL;DR
Symbolic chord recognition remains underexplored because accurate annotations are scarce, and existing methods often overlook human analytical practices. The paper introduces POP909-CL and BACHI, which combines boundary detection with confidence-ordered iterative prediction of chord elements. BACHI achieves state-of-the-art performance on classical and pop benchmarks, while confidence-ordered decoding reflects repertoire-specific patterns.
Problem
Symbolic chord recognition has limited accurate annotated data and methods often overlook alignment with human annotation practices.
Source: §1 · p. 1Method
The paper introduces POP909-CL with human-corrected labels and BACHI, which detects chord boundaries and iteratively predicts chord root, quality, and bass in confidence order.
Source: §1 · p. 1Results
Takeaways & Limitations
The authors conclude that learned confidence patterns offer insights into repertoire-specific harmonic tendencies and that the framework provides a foundation for chord-conditioned music generation.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Automatic chord recognition (ACR) via deep learning models has gradually achieved promising recognition accuracy, yet two key challenges remain. First, prior work has primarily focused on audio-domain ACR, while symbolic music (e.g., score) ACR has received limited attention due to data scarcity. Second, existing methods still overlook strategies that are aligned with human music analytical practices. To address these challenges, we make two contributions: (1) we introduce POP909-CL, an enhanced version of POP909 dataset with tempo-aligned content and human-corrected labels of chords, beats, keys, and time signatures; and (2) We propose BACHI, a symbolic chord recognition model that decomposes the task into different decision steps, namely boundary detection and iterative ranking of chord root, quality, and bass (inversion). This mechanism mirrors the human ear-training practices. Experiments demonstrate that BACHI achieves state-of-the-art chord recognition performance on both classical and pop music benchmarks, with ablation studies validating the effectiveness of each module.
1. INTRODUCTION
Automatic chord recognition annotates music with chord labels over time and supports applications including harmonic analysis, controllable music generation, and music education. Symbolic ACR faces data scarcity and methods less aligned with human analytical practices; the paper introduces POP909-CL and BACHI to address these challenges.
1. INTRODUCTION
- Symbolic ACR remains underexplored amid data scarcity: accurate symbolic chord labels are limited, and existing classical corpora often contain duplicates or short excerpts.Prior approaches consequently relied on subsets or combined chunks for training.Source: §1 · p. 1
- Prior symbolic ACR methods pay little attention to aligning chord recognition with human annotation practices, despite diverse rule-based and neural approaches.The paper identifies transferring human annotation logic as a promising direction.Source: §1 · p. 1
- BACHI decomposes recognition into boundary detection and confidence-ordered iterative prediction of chord root, quality, and bass, reflecting human ear-training practices.Figure 1 depicts the backbone, boundary detection and conditioning, and iterative decoding stages.Source: §1 · p. 1, Figure 1 · p. 2
- POP909-CL enhances POP909 with human-corrected chord, beat, key, and time-signature labels as a resource for MIR research.The enhanced labels address data scarcity in symbolic chord recognition.Source: §1 · p. 1
- BACHI achieves state-of-the-art chord recognition performance on both classical and pop music benchmarks, with ablations validating each module.The paper also presents POP909-CL as a resource for symbolic MIR research.Source: §1 · p. 1
2. METHOD
BACHI represents score input as piano rolls and chord labels as root, quality, and bass. It detects chord boundaries, then decodes chord elements iteratively using masked transformers and confidence ranking.
2.1. Input and Output
- BACHI encodes scores as piano rolls and represents each chord label with root, quality, and bass, aligned to the input at T/6 resolution.The piano rolls use 88 pitch dimensions and 12 frames per beat; a patch embedding maps them to transformer tokens.Source: §2.1 · p. 2
2.2. Two-stage Chord Recognition
- The first recognition stage predicts chord boundaries from encoder states and conditions subsequent chord recognition on them through FiLM.The boundary labels are binarized from chord labels; FiLM uses scale γ and bias β to condition the latent representation.Source: §2.2 · p. 2
- The decoder uses a masked-transformer paradigm, combining self-attention over chord elements with cross-attention to local context.During training, randomly masked chord elements are filled using information from existing elements and the local context.Source: §2.2 · p. 2
- At inference, BACHI ranks confidence across unfilled chord elements, commits the highest-confidence prediction, and repeats until all three are filled.The decoder predicts masked fields simultaneously rather than autoregressively; the resulting order-agnostic procedure is described as mirroring human ear-training practices.Source: §2.2 · p. 2
3. EXPERIMENTS
The experiments construct classical and pop training sets, prepare POP909-CL with corrected annotations, and specify evaluation baselines and ablation variants.
3.1. Dataset and Training Setup
- The classical corpus combines de-duplicated When-in-Rome and DCML annotations, converted from functional harmony to absolute chord labels.The conversion uses music21 and a self-written script for chord quality.Source: §3.1 · p. 3
- POP909-CL provides corrected chord, beat, key, and time-signature labels across all 909 tracks, addressing errors and tempo variation in original POP909.Musicians reviewed the original rule-based labels; the reported corrections include 40.6% misaligned start beats, 14.2% missing key-signature changes, and 2.6% incorrect time signatures.Source: §3.1 · p. 3, Figure 2 · p. 3
- Training uses 1,500 classical and 909 pop pieces, separate genre-specific models, 9:1 train-test splits, and 12-key augmentation.POP909 samples are aligned to a fixed tempo using corrected beat labels; test sets use canonical versions.Source: §3.1 · p. 3
3.2. Evaluation, Ablations and Baselines
- Evaluation measures macro-accuracy for root, quality, and bass, plus overall chord accuracy, against three retrained neural baselines and a rule-based lower-bound anchor.Harmony Transformer v2 is adapted to output the same three chord targets.Source: §3.2 · p. 3
- Ablations compare the full BACHI model with versions lacking boundary detection and iterative decoding, lacking iterative decoding alone, or adding key detection.Classical ablation results appear in the main text, while POP909-CL results and full MIREX-style comparisons are provided on the demo website.Source: §3.2 · p. 3
4. RESULTS
BACHI performs strongly on both classical and pop benchmarks, though classical chord recognition remains difficult. Ablations indicate that boundary detection and iterative decoding work together to improve recognition, while extra key conditioning can hurt.
4.1. Comprehensive Performance
- BACHI leads classical evaluation at 68% full chord accuracy; on pop it ranks first in full-chord and quality accuracy, and second in root and bass accuracy.All classical models remain below 70% full chord accuracy, while most pop models exceed 75%.Source: §4.1 · pp. 3–4
- POP909-CL corrects approximately 35% of POP909 chord-label errors; pop errors cluster among related qualities, while classical errors are more broadly distributed.The reported patterns characterize pop harmony as more concentrated and classical harmony as more variable and annotation-ambiguous.Source: §4.1 · p. 4
4.2. Ablation Study
- 66.8% full chord accuracy is achieved by the basic transformer without boundary detection or iterative decoding, alongside 78.0% root, 78.9% quality, and 77.3% bass accuracy.The baseline removes both components and has lower full-chord accuracy despite reasonable individual-component scores.Source: §4.2 · p. 4
- Combining boundary detection and iterative decoding substantially enhances recognition; key conditioning slightly lowers full-chord accuracy, likely through propagated key-prediction errors.Boundary detection primarily reduces jitter, most evidently on pop, while iterative decoding contributes to overall accuracy.Source: §4.2 · p. 4
5. DISCUSSION
Confidence-guided decoding follows different chord-element orders across classical and pop repertoires. Classical predictions most often begin with quality, while POP909-CL predictions most often begin with bass.
5. DISCUSSION
- Classical predictions begin with quality 40.8% of the time, most often following quality→root→bass (33.2%); POP909-CL predictions begin with bass 66.9%, most often following bass→root→quality (56.4%).The reported orders are consistent with voice-leading cues in classical analysis and bass-led cues in pop.Source: §5 · p. 4
- The confusion matrices support the paper's claim that human-mimicking decision paths benefit symbolic chord recognition over fixed-order decoding.The caption identifies chord-quality confusion matrices for the classical corpus and POP909-CL evaluation sets.Source: Figure 3 · p. 4
6. CONCLUSION
The conclusion presents BACHI as a boundary-aware transformer with confidence-guided masked decoding, alongside human-annotated POP909-CL chord labels. It links these resources to music generation and repertoire-specific harmonic analysis.
6. CONCLUSION
- BACHI combines patch embedding, FiLM-based boundary conditioning, and confidence-guided masked-filling decoding for symbolic chord recognition.Its confidence-ordered decoding favors quality-first prediction in classical music and bass-first prediction in popular music.Source: §6 · p. 4
- The framework and POP909-CL annotations are presented as a foundation for harmonically coherent, structurally aware generation and for studying repertoire-specific harmonic tendencies.The paper reports that code, trained models, and annotations are released on its demo page.Source: §6 · p. 4