Score-Agnostic Structure Analysis in Large-Scale Performance Datasets

Gerhard Widmer; Patricia Hu; Silvan Peter

REVIEW 2 major objections 1 minor 13 references

Transcriptions of the same piano piece can be grouped by structural realizations using alignment costs and sequence length differences without any score reference.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-29 20:22 UTC pith:KYIABCGN

load-bearing objection The paper outlines a pipeline to cluster transcriptions by structure using alignment cost and length dissimilarity, but the abstract contains no results or checks to show it works. the 2 major comments →

arxiv 2605.25951 v1 pith:KYIABCGN submitted 2026-05-25 cs.SD

Score-Agnostic Structure Analysis in Large-Scale Performance Datasets

Patricia Hu , Silvan Peter , Gerhard Widmer This is my paper

classification cs.SD

keywords automatic music transcriptionstructural analysissequence alignmenthierarchical clusteringpiano performancescore-agnostic methodsperformance datasets

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper demonstrates a method to organize large collections of automatically transcribed piano performances according to their underlying musical structures. Classical performances often differ in repeats, cuts, or edition-specific choices, so mixing them invalidates direct comparisons of expressive features. Pairwise sequence-to-sequence alignments are computed for every pair of transcriptions of one piece, and the resulting alignment costs together with differences in performed sequence lengths serve as features for hierarchical clustering. Clusters that share the same structural pattern then support valid within-group performance analysis on datasets that lack ground-truth scores or audio. The approach is shown on roughly 1500 transcriptions covering 88 compositions.

Core claim

We address this by applying sequence-to-sequence alignment followed by hierarchical clustering: we create pairwise alignments for all pairs of transcriptions of a given piece, and use the alignment cost and (dis)similarity of performed sequence lengths to resolve structural mismatches as features for grouping. We propose this approach as a first step towards automatically evaluating large-scale transcribed datasets that lack ground-truth score and/or audio, shifting the evaluation criterion from truth-based accuracy to musical coherence and plausibility.

What carries the argument

Sequence-to-sequence alignment followed by hierarchical clustering, using alignment cost and performed sequence length dissimilarity as features to group transcriptions that share the same structural realization.

Load-bearing premise

Alignment costs combined with differences in performed sequence lengths are sufficient to separate distinct structural realizations into musically meaningful clusters.

What would settle it

Check whether the clusters produced on a subset of transcriptions that also have known ground-truth scores match the actual structural variants such as different repeat patterns or edition choices.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Transcriptions within each cluster share a common structural interpretation and can be compared directly for expressive features.
Large transcribed datasets can be partitioned into structurally coherent subsets without reference to the original score.
Evaluation of automatic transcriptions can be based on musical coherence within clusters rather than exact match to a reference.
Structural choices across many performances of the same piece become quantifiable at scale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The clusters could later be used to study how often particular structural variants occur across different performers.
Within each structural group the same alignment machinery might help detect local transcription errors.
Extending the method to other ensemble types could reveal whether structural coherence behaves similarly outside solo piano.
If clusters prove stable across different alignment algorithms, the approach could serve as a preprocessing step for any performance corpus.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

The paper outlines a pipeline to cluster transcriptions by structure using alignment cost and length dissimilarity, but the abstract contains no results or checks to show it works.

read the letter

The main takeaway is a straightforward pipeline: align every pair of transcriptions for a piece, then cluster on alignment cost plus performed length difference to separate structural variants like repeats or edition differences. They run this on roughly 1500 transcriptions of 88 pieces from a large AMT dataset.

What is actually new is the targeted use of those two scalars for score-free grouping in transcribed performance collections. Standard alignment methods exist, but applying them this way to organize AMT output for later performance comparison is a direct and relevant step.

The paper does a clear job naming the practical problem. Large transcribed sets mix expressive differences with structural ones, and you need some way to sort them before studying tempo or dynamics across versions.

The soft spot is the total lack of evidence. The abstract describes the steps but reports no cluster quality numbers, no manual inspection of groups, no comparison against pieces with known scores, and no test for whether the features separate structure from tempo, ornaments, or transcription noise. That leaves the central assumption untested.

The stress-test concern lands: alignment cost reacts to many local changes, and length difference alone does not locate where the structural edits occur. Without normalization details or consistency checks, it is unclear whether the clusters end up musically coherent.

This is for people in computational musicology or MIR who clean or organize large transcribed piano datasets. A reader facing that exact task might borrow the alignment-plus-clustering idea as a starting point.

It deserves peer review because the problem is real and the method is simple enough to implement, but the authors will need to add concrete validation before the claim holds.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a score-agnostic method to group automatic transcriptions of the same classical piano piece by structural realization (repeat patterns, edition variants). It computes pairwise sequence-to-sequence alignments for all transcriptions of a piece and feeds the alignment cost together with performed-sequence-length dissimilarity into hierarchical clustering. The approach is presented as a first step for evaluating large AMT datasets without ground-truth scores or audio and is demonstrated on approximately 1,500 transcriptions of 88 compositions.

Significance. If the two scalar features reliably separate structural variants from other sources of variation, the method would allow performance researchers to work with large transcribed corpora that currently lack usable reference material. The paper correctly identifies a real practical bottleneck, but the absence of any quantitative validation or internal consistency checks means the significance cannot yet be assessed beyond the level of a plausible procedural idea.

major comments (2)

[Demonstration] Demonstration paragraph: the claim that the method groups transcriptions 'according to their underlying structural realisation' is unsupported because no quantitative results (cluster validity indices, manual inspection of cluster contents, or comparison against known structural variants) are reported for the 1,500-transcription experiment.
[Method] Method description: the assertion that alignment cost and length dissimilarity 'resolve structural mismatches as features for grouping' does not address the known sensitivity of alignment cost to local timing deviations, ornaments, and transcription substitutions, nor does it explain how length dissimilarity distinguishes inserted repeats from omitted sections.

minor comments (1)

[Abstract] The abstract states that the evaluation criterion shifts to 'musical coherence and plausibility' but provides no operational definition or procedure for assessing coherence in the demonstration.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important areas where the manuscript can be strengthened, particularly in providing quantitative support for the demonstration and expanding the method description. We address each major comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [Demonstration] Demonstration paragraph: the claim that the method groups transcriptions 'according to their underlying structural realisation' is unsupported because no quantitative results (cluster validity indices, manual inspection of cluster contents, or comparison against known structural variants) are reported for the 1,500-transcription experiment.

Authors: We agree that the current demonstration lacks explicit quantitative validation and that this weakens the claim. In the revised manuscript we will add cluster validity indices (e.g., silhouette coefficient and Davies-Bouldin index) computed on the hierarchical clustering results, include a summary table of cluster sizes and purity for a representative subset of the 88 pieces, and report any available manual checks against known repeat patterns or edition variants present in the source dataset. revision: yes
Referee: [Method] Method description: the assertion that alignment cost and length dissimilarity 'resolve structural mismatches as features for grouping' does not address the known sensitivity of alignment cost to local timing deviations, ornaments, and transcription substitutions, nor does it explain how length dissimilarity distinguishes inserted repeats from omitted sections.

Authors: The referee is correct that the method section is too terse on these issues. We will revise it to (1) acknowledge the sensitivity of alignment cost to local timing, ornaments and substitutions, (2) explain that the hierarchical clustering operates on the joint distribution of alignment cost and length dissimilarity so that global structural differences dominate over local noise, and (3) clarify that length dissimilarity is computed after optimal alignment and therefore flags large-scale insertions or omissions that cannot be explained by local timing variation alone. revision: yes

Circularity Check

0 steps flagged

No circularity; purely procedural method with no self-referential reductions

full rationale

The paper presents a direct procedural pipeline of pairwise sequence alignment followed by hierarchical clustering that uses alignment cost and performed-sequence-length dissimilarity as input features. No equations, fitted parameters, predictions, or self-citations are described that would reduce any claimed result to its own inputs by construction. The approach is offered as an application of standard alignment and clustering techniques rather than a derivation whose validity depends on prior self-referential steps. This is the normal case of a self-contained methodological description.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that the chosen features capture structural mismatches without ground truth; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Alignment cost and sequence length dissimilarity are sufficient features to distinguish structural realizations
Invoked when stating these are used as features for grouping transcriptions.

pith-pipeline@v0.9.1-grok · 6289 in / 1172 out tokens · 48510 ms · 2026-06-29T20:22:06.630340+00:00 · methodology

0 comments

read the original abstract

In recent years, thanks to advances in automatic music transcription (AMT), several large-scale datasets of automatically transcribed piano solo music have been released. While these datasets undoubtedly offer extensive material for performance studies, they vary substantially in quality. In the case of classical music, performances often differ not only in expressive aspects such as tempo, but also in their structural interpretation of the score (including repeat patterns and edition-specific variants). To meaningfully use large-scale transcribed datasets for performance research, transcriptions of the same piece must be grouped according to their underlying structural realisation to support valid comparison. We address this by applying sequence-to-sequence alignment followed by hierarchical clustering: we create pairwise alignments for all pairs of transcriptions of a given piece, and use the alignment cost and (dis)similarity of performed sequence lengths to resolve structural mismatches as features for grouping. We propose this approach as a first step towards automatically evaluating large-scale transcribed datasets that lack ground-truth score and/or audio, shifting the evaluation criterion from truth-based accuracy to musical coherence and plausibility. We demonstrate our score-agnostic approach on around 1,500 transcriptions of 88 compositions from a recently published large-scale transcribed piano performance dataset.

Figures

Figures reproduced from arXiv: 2605.25951 by Gerhard Widmer, Patricia Hu, Silvan Peter.

**Figure 1.** Figure 1: Four distance matrices derived from all pairwise alignments of transcriptions from the 3rd movement of the Piano [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** Figure 2: A comparison of the estimated grouping by our score-agnostic clustering (left) and the true (manually verified) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: A dendrogram visualisation of the hierarchical clustering process as a tree-like diagram (same piece and tran [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** Figure 4: Score image excerpt taken from the second movement of the Kreisleriana Op. 16 by R. Schumann from the [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages · 2 internal anchors

[1]

To date, most work in expressive performance analysis has relied on small, curated datasets and/or manual annotation

INTRODUCTION Music performance research involves understanding and interpreting the nuances of expressive musical renditions. To date, most work in expressive performance analysis has relied on small, curated datasets and/or manual annotation. In the case of solo piano music, most work has relied on MIDI-like data measured on computer-controllable instru-...
[2]

Score-Agnostic Structure Analysis in Large-Scale Performance Datasets

METHOD Our proposed method works in two steps: first, we com- pute alignments of all pairs of transcriptions of a given piece using Dynamic Time Warping (DTW) with a cus- tom distance metric. Second, we cluster the transcriptions based on their alignment cost, stretch (warping) and se- quence length (dis)similarity. Copyright: © 2026. This is an open-acce...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

DEMONSTRA TION We demonstrate our method on the ATEPP dataset [3], which contains 11,674 transcriptions of performances by world-renowned pianists of 1,595 unique compositions, for 319 (20%) of which score files are provided. We test our method on all transcriptions of works by Haydn, Mozart, Beethoven, Schubert and Schumann which pro- vide a correspondin...
[4]

CONCLUSION We presented an approach to automatically align and clus- ter transcriptions of a given composition into their differ- ent structural score realisations. We propose this approach as a first step towards automatically evaluating transcribed datasets that lack ground-truth score and/or audio, shifting the evaluation criterion from truth-based acc...

2020
[5]

High- resolution piano transcription with pedals by regress- ing onset and offset times,

Q. Kong, B. Li, X. Song, Y . Wan, and Y . Wang, “High- resolution piano transcription with pedals by regress- ing onset and offset times,”IEEE/ACM Transactions on Audio Speech and Language Processing, vol. 29, pp. 3707–3717, 2021

2021
[6]

GiantMIDI- Piano: A large-scale midi dataset for classical piano music,

Q. Kong, B. Li, J. Chen, and Y . Wang, “GiantMIDI- Piano: A large-scale midi dataset for classical piano music,”Transactions of the International Society for Music Information Retrieval, 2020

2020
[7]

ATEPP: A dataset of automat- ically transcribed expressive piano performance,

H. Zhang, J. Tang, S. Rafee, S. Dixon, G. Fazekas, G. Wigginset al., “ATEPP: A dataset of automat- ically transcribed expressive piano performance,” in Proceedings of the 24th International Society for Mu- sic Information Retrieval Conference, ISMIR, 2023

2023
[8]

Aria-MIDI: A dataset of piano midi files for symbolic music modeling,

L. Bradshaw and S. Colton, “Aria-MIDI: A dataset of piano midi files for symbolic music modeling,” inIn- ternational Conference on Learning Representations (ICLR), 2025

2025
[9]

Dynamic programming algo- rithm optimization for spoken word recognition,

H. Sakoe and S. Chiba, “Dynamic programming algo- rithm optimization for spoken word recognition,”IEEE transactions on acoustics, speech, and signal process- ing, vol. 26, no. 1, pp. 43–49, 1978

1978
[10]

Automatic note-level score-to-performance align- ments in the asap dataset,

S. D. Peter, C. E. Cancino-Chac ´on, F. Foscarin, A. P. McLeod, F. Henkel, E. Karystinaios, and G. Widmer, “Automatic note-level score-to-performance align- ments in the asap dataset,”Transactions of the Interna- tional Society for Music Information Retrieval, vol. 6, no. 1, 2023

2023
[11]

Modern hierarchical, agglomerative clustering algorithms

D. M ¨ullner, “Modern hierarchical, agglomerative clus- tering algorithms,”arXiv preprint arXiv:1109.2378, 2011

work page internal anchor Pith review Pith/arXiv arXiv 2011
[12]

How to infer re- peat structures in midi performances,

S. Peter, P. Hu, and G. Widmer, “How to infer re- peat structures in midi performances,”arXiv preprint arXiv:2505.05055, 2025

work page arXiv 2025
[13]

V-measure: A con- ditional entropy-based external cluster evaluation mea- sure,

A. Rosenberg and J. Hirschberg, “V-measure: A con- ditional entropy-based external cluster evaluation mea- sure,” inProceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP- CoNLL), 2007, pp. 410–420. Figure 4. Score image excerpt taken from the second movement of the K...

2007

[1] [1]

To date, most work in expressive performance analysis has relied on small, curated datasets and/or manual annotation

INTRODUCTION Music performance research involves understanding and interpreting the nuances of expressive musical renditions. To date, most work in expressive performance analysis has relied on small, curated datasets and/or manual annotation. In the case of solo piano music, most work has relied on MIDI-like data measured on computer-controllable instru-...

[2] [2]

Score-Agnostic Structure Analysis in Large-Scale Performance Datasets

METHOD Our proposed method works in two steps: first, we com- pute alignments of all pairs of transcriptions of a given piece using Dynamic Time Warping (DTW) with a cus- tom distance metric. Second, we cluster the transcriptions based on their alignment cost, stretch (warping) and se- quence length (dis)similarity. Copyright: © 2026. This is an open-acce...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[3] [3]

DEMONSTRA TION We demonstrate our method on the ATEPP dataset [3], which contains 11,674 transcriptions of performances by world-renowned pianists of 1,595 unique compositions, for 319 (20%) of which score files are provided. We test our method on all transcriptions of works by Haydn, Mozart, Beethoven, Schubert and Schumann which pro- vide a correspondin...

[4] [4]

CONCLUSION We presented an approach to automatically align and clus- ter transcriptions of a given composition into their differ- ent structural score realisations. We propose this approach as a first step towards automatically evaluating transcribed datasets that lack ground-truth score and/or audio, shifting the evaluation criterion from truth-based acc...

2020

[5] [5]

High- resolution piano transcription with pedals by regress- ing onset and offset times,

Q. Kong, B. Li, X. Song, Y . Wan, and Y . Wang, “High- resolution piano transcription with pedals by regress- ing onset and offset times,”IEEE/ACM Transactions on Audio Speech and Language Processing, vol. 29, pp. 3707–3717, 2021

2021

[6] [6]

GiantMIDI- Piano: A large-scale midi dataset for classical piano music,

Q. Kong, B. Li, J. Chen, and Y . Wang, “GiantMIDI- Piano: A large-scale midi dataset for classical piano music,”Transactions of the International Society for Music Information Retrieval, 2020

2020

[7] [7]

ATEPP: A dataset of automat- ically transcribed expressive piano performance,

H. Zhang, J. Tang, S. Rafee, S. Dixon, G. Fazekas, G. Wigginset al., “ATEPP: A dataset of automat- ically transcribed expressive piano performance,” in Proceedings of the 24th International Society for Mu- sic Information Retrieval Conference, ISMIR, 2023

2023

[8] [8]

Aria-MIDI: A dataset of piano midi files for symbolic music modeling,

L. Bradshaw and S. Colton, “Aria-MIDI: A dataset of piano midi files for symbolic music modeling,” inIn- ternational Conference on Learning Representations (ICLR), 2025

2025

[9] [9]

Dynamic programming algo- rithm optimization for spoken word recognition,

H. Sakoe and S. Chiba, “Dynamic programming algo- rithm optimization for spoken word recognition,”IEEE transactions on acoustics, speech, and signal process- ing, vol. 26, no. 1, pp. 43–49, 1978

1978

[10] [10]

Automatic note-level score-to-performance align- ments in the asap dataset,

S. D. Peter, C. E. Cancino-Chac ´on, F. Foscarin, A. P. McLeod, F. Henkel, E. Karystinaios, and G. Widmer, “Automatic note-level score-to-performance align- ments in the asap dataset,”Transactions of the Interna- tional Society for Music Information Retrieval, vol. 6, no. 1, 2023

2023

[11] [11]

Modern hierarchical, agglomerative clustering algorithms

D. M ¨ullner, “Modern hierarchical, agglomerative clus- tering algorithms,”arXiv preprint arXiv:1109.2378, 2011

work page internal anchor Pith review Pith/arXiv arXiv 2011

[12] [12]

How to infer re- peat structures in midi performances,

S. Peter, P. Hu, and G. Widmer, “How to infer re- peat structures in midi performances,”arXiv preprint arXiv:2505.05055, 2025

work page arXiv 2025

[13] [13]

V-measure: A con- ditional entropy-based external cluster evaluation mea- sure,

A. Rosenberg and J. Hirschberg, “V-measure: A con- ditional entropy-based external cluster evaluation mea- sure,” inProceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP- CoNLL), 2007, pp. 410–420. Figure 4. Score image excerpt taken from the second movement of the K...

2007