Source-linked AI summary
Local chord corruption is not recognizer replay: structure-matched calibration
Weiwen Huang, Yunda Chen, Wangzheng Wu, Nengheng Zheng
TL;DR
Synthetic chord substitutions provide controlled tests, but may not reproduce the effects of complete recognized chord sequences. The paper matches synthetic profiles to replay structure and evaluates response magnitude alongside output-chord agreement. Structure matching narrows response-distance gaps to replay, while joint matching better reproduces replay-generated chords.
Problem
A local chord substitution may not reproduce a generator’s response to a complete recognized sequence, so representativeness must be tested separately from sensitivity to an isolated change.
Source: §1 · p. 1Method
Structure-matched calibration resamples chord labels while matching recognizer replay’s changed positions and harmonic-relation composition, then compares paired generations on response and output-chord agreement.
Source: §1 · p. 1, §2.1 · p. 2Results
On MoisesDB, structure matching reduced CENS response distance by 81% for MIDI-SAG and 77% for MusicGen-Chord, improving every song in both models.
Source: §3.2 · p. 3Takeaways & Limitations
Recognized harmony should be evaluated as a sequence, reporting both response magnitude and output structure; joint matching best reproduces replay-generated chords.
Source: §4 · p. 4Replay fidelity remains untested in generators without explicit chord controls and is sensitive to temporal alignment.
Source: §3.4 · p. 4
Abstract
from arXiv · showhide
Synthetic chord substitutions offer controlled tests of music generation, but their effects can differ from those of a complete recognized chord sequence. We propose structure-matched calibration, which constructs synthetic chord sequences that match the changed positions and harmonic-relation composition of recognizer replay. Paired generation compares both target-response magnitude and output-chord agreement with replay. On 29 of 30 MUSDB18-HQ songs, central four-second tritone corruption produces a larger target response than complete replay in MIDI-SAG. On 24 held-out MoisesDB songs, structure matching reduces target-response distance to replay by 81% for MIDI-SAG and 77% for MusicGen-Chord. Distance decreases on every song in both models with CNN--CRF. Joint matching also reduces output-chord mismatch with replay by 8--17 percentage points relative to temporal or relational matching alone.
1. INTRODUCTION
Chord-conditioned accompaniment makes harmony an important control, but a local synthetic substitution may not reproduce the effect of a complete recognized chord sequence. Structure-matched calibration tests that representativeness; results show local and replay conditions can differ, while matching replay structure narrows the gap.
1. INTRODUCTION
- Chord sequences are an explicit harmony control in recent singing-accompaniment systems, so recognition differences can propagate into generated accompaniment.The section frames understanding that propagation as important for evaluating chord-conditioned systems.Source: §1 · p. 1
- Prior synthetic substitutions and adherence or coherence measures assess different questions from whether a probe reproduces the downstream effect of a complete recognized sequence.Related chord-recognition work also examines label vocabularies, timing boundaries, harmonic relations, pitch-space geometry, and broader ACR evaluation.Source: §1 · p. 1
- Structure-matched calibration resamples chord labels while matching replay’s changed positions and harmonic-relation composition, and was tested across three generators, two recognizers, and two datasets.The generators were MIDI-SAG, MusicGen-Chord, and AccoMontage; the recognizers were CNN–CRF and DeepChroma+CRF.Source: §1 · p. 1
- Structure matching reduced target-response distance to replay by 81% for MIDI-SAG and 77% for MusicGen-Chord on 24 held-out MoisesDB songs, improving every song in both models.The comparison used CNN–CRF replay.Source: §1 · p. 1
- On 29 of 30 MUSDB18-HQ songs, a four-second central tritone probe produced a larger target response than complete replay in MIDI-SAG.Replay divergence from the reference spanned 10.73 seconds on average, more than twice the probe duration.Source: §1 · p. 1
2. METHOD
The method pairs chord conditions while holding other generation inputs fixed, constructs profiles that match replay structure, and evaluates both target response and generated-chord agreement. Its protocol applies paired comparisons across generators, datasets, and recognizers.
2.1. Experimental setup
- Paired generation holds the generator, non-chord inputs, and seed fixed while comparing baseline, full-mixture replay, a central four-second tritone corruption, and a replay-matched profile.Replay comes from full-mixture ACR; the reference comes from vocal-track harmonization, and the profile may coincide with replay where a category has no alternative label.Source: §2.1 · pp. 1–2, Figure 1 · p. 2
2.2. Profile construction
- Profiles preserve replay’s changed positions and relation-category composition; MoisesDB additionally matches the relation category at each changed position.Labels are sampled within matched categories, preferring alternatives to replay where possible, while baseline labels remain elsewhere.Source: §2.2 · p. 2
2.3. Evaluation metrics
- Target response compares altered-output chroma with the baseline and altered chord templates over changed blocks; positive values indicate movement toward the altered condition’s chord target.Response is averaged over three generation seeds, and calibration gain measures whether the profile is closer to replay than the central corruption; CENS is primary and CQT is a second representation.Source: §2.3 · p. 2
- Output-chord mismatch compares chord sequences decoded from generated waveforms, separately testing whether a condition reproduces replay’s harmonic consequences.The generated waveforms, including replay, are decoded with DeepChroma+CRF on the same 24-cell grid.Source: §2.3 · p. 3
2.4. Experimental protocol
- The protocol scores 30 MUSDB18-HQ songs and 24 MoisesDB songs, tests temporal-only, relation-only, and joint matching, and uses song-level bootstrap intervals and paired signed-rank tests.MIDI-SAG and MusicGen-Chord use three paired seeds per condition; statistical comparisons use 10,000 bootstrap resamples and specified multiple-comparison corrections.Source: §2.4 · p. 3
3. RESULTS
Structure-matched profiles brought generated responses closer to recognizer replay across models, while local tritone corruption often produced a larger response than replay. Matching both temporal and relational structure also improved output-chord agreement, though response magnitude and harmonic consequences remain separate evaluation targets.
3.1. Local corruption and recognizer replay
- On 29 of 30 MUSDB18-HQ songs, the four-second central tritone probe produced a larger target response than complete CNN–CRF replay.Replay divergence spanned 10.73 seconds on average and was dominated by root changes; the figure marks the single reversal.Source: §3.1 · p. 3, Figure 2 · p. 3
- 0.482 to 0.098: MIDI-SAG’s mean CENS distance to replay fell with profile calibration on MUSDB18-HQ, with the improvement also holding for a second recognizer and chroma representation.The mean-distance difference was 0.384.Source: §3.1 · p. 3, Figure 3 · p. 4
3.2. Cross-model results
- 81% for MIDI-SAG and 77% for MusicGen-Chord: structure matching reduced CENS response distance on MoisesDB, improving all 24 songs under CNN–CRF.The reduction persisted with DeepChroma+CRF and when replay distances were averaged over three independently sampled profiles.Source: §3.2 · p. 3, Table 1 · p. 3
- Approximately 74%: calibration reduced AccoMontage’s target-response distance on 23 of 24 songs, transferring to its native beat-based interface.The target-response and raw pitch-class tests gave p = 2.98 × 10−6 and p = 2.47 × 10−5, respectively.Source: §3.2 · p. 4, Table 2 · p. 4
3.3. Ablation study
- Joint matching had the lowest mean CENS distance, but none of the four full-versus-single-factor comparisons was significant after Holm correction.Both single-factor conditions reduced mean CENS distance; the figure separates this measure from decoded output-chord mismatch.Source: §3.3 · p. 4, Figure 4 · p. 4
- 14.35–16.96 percentage points for MIDI-SAG and 8.04–11.05 for MusicGen-Chord: joint matching reduced output-chord mismatch versus either single-factor condition.All four comparisons remained significant after Holm correction, and the full profile was favored on 17 to 21 of 24 songs.Source: §3.2 · pp. 3–4, §3.3 · p. 4
3.4. Discussion
- Synthetic perturbations should match the harmony-control interface; replay fidelity is sensitive to temporal alignment and remains untested in generators without explicit chord controls.The discussion proposes testing structural agreement alongside perceived accompaniment quality in a recognizer-in-the-loop benchmark.Source: §3.4 · p. 4
4. CONCLUSION
Structure-matched profiles narrow the gap between local corruption and recognizer replay across MIDI-SAG, MusicGen-Chord, and AccoMontage. The conclusion recommends evaluating recognized harmony as a sequence using both response magnitude and output structure.
4. CONCLUSION
- Structure-matched profiles narrow the gap between local corruption and recognizer replay across MIDI-SAG, MusicGen-Chord, and AccoMontage.On 24 MoisesDB songs, both audio models improve on every song; repeated profiles and a second recognizer preserve the effect.Source: §4 · p. 4
- Joint matching best reproduces replay-generated chords.It reduces mismatch more than temporal or relational matching alone.Source: §4 · p. 4
- Recognized harmony should be evaluated as a sequence, reporting response magnitude and output structure together.Source: §4 · p. 4