Pith. sign in

REVIEW 4 major objections 5 minor 42 references

A model using only performance features — timing, dynamics, articulation — recovers the direction of emotion differences between performances of the same piece, with a mean pairwise angular error of about 8.4 degrees, though it underestimat

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 · deepseek-v4-flash

2026-08-03 01:25 UTC pith:5UURXH3R

load-bearing objection Useful relative emotion prediction with a real directional signal, but the evidence chain has two missing links: transcription velocity robustness and a zero-deviation baseline. the 4 major comments →

arxiv 2607.28876 v1 pith:5UURXH3R submitted 2026-07-30 cs.SD cs.LGcs.MM

Learning to Predict Performance-induced Emotion Differences in Classical Piano Music

Joann Ching , Gerhard Widmer This is my paper
classification cs.SD cs.LGcs.MM
keywords music emotion recognitionvalence-arousalexpressive performanceperformance codecautomatic piano transcriptionrelative regressionpairwise difference predictionclassical piano
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.

This paper claims that the emotional differences between performances of the same piece can be predicted from performance-only features, without any input about the composition itself. The Delta-VA model predicts each performance's deviation in valence and arousal from the piece's average annotation using a note-level performance codec (beat period, velocity, timing, articulation). Evaluated on pairwise differences between performances, predicted difference vectors align with ground truth at a mean angular error of about 8.4 degrees — high directional consistency — but with a mean magnitude ratio near 0.48, meaning the model systematically underestimates how large the affective shift is. The authors argue this directional signal is the practically useful part, supporting applications like recommending an alternative performance that feels more positive or calmer.

Core claim

The central claim is that the direction of performance-induced emotion change is recoverable from expressive performance features alone. Delta-VA learns to output a deviation vector (ΔValence, ΔArousal) relative to the piece's 'score-state' — the centroid of all human annotations for that piece. When the evaluation shifts from score-state-anchored absolute error to pairwise difference vectors between two performances of the same piece, the model achieves a mean angular error of 8.374° ± 10.552 (with 0° being perfect) and a mean magnitude ratio of 0.478 ± 0.184, indicating systematic compression. The high R² of 0.786 on the absolute task is, the authors argue, misleading because it depends on

What carries the argument

The Performance Codec — four note-level parameters (beat period, MIDI velocity, timing deviation, articulation ratio) that describe a performance relative to the score, drawn from the expressive-rendering literature. Extracted from commercial audio via an automatic transcription model, these features deliberately exclude composition-level information such as key, mode, or pitch. Delta-VA processes each feature block through a dedicated linear projection into a 16-dimensional embedding, concatenates the embeddings, passes them through a shared ReLU layer, and outputs valence and arousal deviations from the score-state. Its evaluation rests on two geometric metrics: Mean Angular Error between

Load-bearing premise

The automatic transcription model estimates MIDI velocities accurately enough to preserve the relative dynamic contrasts between performances — an assumption the paper itself flags as fragile in prior work, supported here only by a comparison table that is omitted for space reasons.

What would settle it

Corrupt only the velocity channel of the performance codec for a held-out piece — e.g., randomly permute velocities across notes or apply a per-performance constant gain while leaving timing, articulation, and beat period intact — and recompute Delta-VA's pairwise Mean Angular Error. If the error stays near 8.4°, the directional signal does not depend on dynamics as the analysis suggests; if it rises sharply, the result hinges on transcription quality, which the paper identifies as fragile.

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

If this is right

  • Perceived emotion in music can be separated into a composition-dominated absolute component and a performance-driven relative component; Delta-VA shows the relative component is learnable from performance features alone.
  • Performance recommendation becomes concrete: queries like 'a more positive, less aggressive version of this piece' reduce to evaluating pairwise difference vectors, which the model preserves in direction.
  • The systematic magnitude compression (MMR = 0.478) implies that for any two performances, the model can reliably tell which one is affectively more positive or arousing, but not by how much — a practical limitation for ranking.
  • The paper's geometric evaluation (pairwise angular error) may be a more informative standard than R² for any regression where relative ordering, not absolute position, is the task.

Where Pith is reading between the lines

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

  • Because velocity shows the highest cross-performer variance and is the most fragile transcription output, the reported direction-alignment may be partly an artifact of velocity estimation; testing on direct-MIDI recordings would separate transcription error from genuine performance signal.
  • The magnitude compression could be reduced by training with a loss that explicitly penalizes magnitude ratio or by adding a calibration head; no such mechanism is currently in the model.
  • The same codec-based relative regression could transfer to non-piano instruments if a reliable transcription and alignment pipeline exists, since the four parameters are score-relative rather than instrument-specific.
  • A stronger test of the paper's utility claim would be a listening experiment: if listeners' actual pairwise preference judgments correlate with the predicted difference-vector directions, the recommendation framing is validated.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Delta-VA, a relative regression framework that predicts performance-induced deviations in valence-arousal from a performance codec (beat period, velocity, timing, articulation) extracted from audio via automatic piano transcription. Using the CP-WTC dataset of six professional recordings of Bach's Well-Tempered Clavier Book I, the authors first show that the codec features vary across performers, then benchmark absolute regressors (linear, kNN, MLP, rank-based) against Delta-VA. Delta-VA adds a per-piece score-state anchor to the predictions and is reported to reach R^2 = 0.786 on the absolute task. The paper then introduces geometric pairwise metrics, reporting a mean angular error of 8.374° ± 10.552 and a mean magnitude ratio of 0.478 ± 0.184, which are interpreted as evidence that the model preserves the direction, though not the magnitude, of perceived affective differences between performances.

Significance. If the claims hold, the paper makes a useful contribution by shifting MER from absolute emotion prediction to performance-conditioned difference prediction, and by proposing evaluation metrics (MAE_theta, MMR) that are more appropriate for relative tasks. The explicit focus on the performance codec and the use of modern transcription models to extract expressive features from commercial recordings is timely. The paper also ships complementary code and uses a public dataset, which supports reproducibility. However, the significance is tempered by the small number of performers (N=6), the omission of the transcription robustness comparison table despite the feature being load-bearing, and the absence of a score-state-only baseline, all of which currently leave the central claim under-supported.

major comments (4)
  1. [Section 3.3] The transcription-model robustness table is omitted 'for space reasons.' This is not merely a presentation issue: Fig. 2b shows velocity has the highest and most consistent cross-performer variance, so the velocity estimates are the dominant performance signal. Section 3.3 itself cites prior work [23,24] showing velocity is fragile under commercial-recording conditions. Without the actual comparison table, the reader cannot assess whether Transkun's velocity estimates are accurate enough to preserve relative dynamics. The variance and RSA analyses in Section 4 use the same potentially corrupted features and therefore cannot independently validate transcription quality. I request that the full comparative results be reported (in an appendix if needed), and that the paper include a direct check of velocity correlation on the CP-WTC recordings or a sensitivity analysis.
  2. [Table 1] The evaluation of Delta-VA lacks a trivial baseline: a model that predicts the score-state anchor µ_score for every performance. Because Delta-VA's absolute prediction is µ_score + y_dev, the reported R^2 = 0.786 mixes the anchor's contribution with the model's deviation predictions. The pairwise analysis in Section 6 partially addresses this by canceling the anchor, but the absolute task is still presented as a headline result. Please add an 'anchor-only' baseline (and, ideally, a 'mean-only' baseline across all pieces) to Table 1, and report R^2 and MSE for those baselines so the reader can separate the gain attributable to the learned deviation model.
  3. [Table 1 / Section 5.1] Table 1 reports only point estimates for MSE and R^2, with no fold-level statistics. With only six performers and an exhaustive 30-fold split, the reported differences between models (e.g., Rank-based vs. Delta-VA) could be driven by a few folds. Since the paper's main conclusion depends on Delta-VA outperforming the baselines, I ask for fold-level means and standard deviations/confidence intervals for all metrics, including the angular error and MMR in Section 6.
  4. [Section 6] The pairwise angular error of 8.374° is remarkably low, but this metric is computed after removing the score-state reference. Because Delta-VA was trained with the anchor in the loss (Eq. 3), the pairwise difference vectors depend on the model's predicted deviations; any systematic shrinkage of y_dev will be partially hidden in the pairwise comparison. The reported MMR of 0.478 already indicates strong magnitude compression. The paper should report the distribution of pairwise angular errors and show how sensitive the 8.374° result is to the anchor removal, e.g., by comparing against a model that always predicts zero deviations (which would give random angular errors) or against the baselines under the same pairwise evaluation protocol.
minor comments (5)
  1. [Abstract] Typo: 'Delta-V A' should be 'Delta-VA' (also appears in Section 5.3).
  2. [Figure 2] Panel (a) labels 'Prae-' instead of 'Prelude' (e.g., 'Prae-CM' vs 'Prelude C major'). Clarify the abbreviation or use full piece names. Also, panel (c) reports RSA similarity values but no significance or interpretation of what range would indicate 'moderate agreement'.
  3. [Section 5.2] The acronym 'RnC' is used without expansion in the text ('we adapt the RnC loss'). It is defined only via reference [32] as 'Rank-n-contrast' in the reference list. Spell out the full name at first use.
  4. [Section 5.3] The notation µ_score is introduced informally as 'the score-state' but never explicitly defined with an equation. Clarify that µ_score is the per-piece mean of the valence-arousal annotations over the training performances, and state whether it is recomputed per fold or fixed across folds. This is relevant to the earlier major comment about the anchor baseline.
  5. [Section 6] Equation (4) uses arccos of a dot product; for numerical stability with near-zero vector norms, define a convention (e.g., set angle to 0 when either vector is zero). Also, the notation for the paired sum in Eq. (4) is slightly ambiguous; clarify that N is the number of performance pairs, not the number of segments.

Circularity Check

0 steps flagged

No constructional circularity: Delta-VA's pairwise metrics are genuine held-out predictions; the missing transcription-robustness table is an evidentiary gap, not a circular step.

full rationale

The derivation chain is self-contained with respect to the target. Delta-VA predicts y_dev = (ΔValence, ΔArousal) relative to a piece-specific score-state μ_score computed from the training corpus, and reconstructs absolute predictions as ŷ = μ_score + y_dev (Eq. 2). In the pairwise evaluation (Eqs. 4–5), μ_score cancels because both true and predicted difference vectors are differences between two performances of the same piece: d_ab = y_a − y_b and d̂_ab = (μ + d̂_a) − (μ + d̂_b) = d̂_a − d̂_b. Thus the reported mean angular error of 8.374° ± 10.552 and MMR of 0.478 are not forced by the anchor; they measure whether held-out feature-to-deviation predictions preserve relational direction. The model is trained by MSE under exhaustive performer-wise splits, so the evaluation is on performers not seen in training. The performance codec [6], dataset [18], and mpteval [30] are self-citations, but they function as tools/data, not as a uniqueness theorem or an ansatz that pre-determines the emotion differences; no equation in the paper defines the prediction target as a function of the fitted parameters. The main weakness is evidentiary: Section 3.3 acknowledges that prior work has shown transcription models, especially velocity estimation, are sensitive to recording conditions [23,24], and then states 'For space reasons, we omit the detailed comparative results table here' before concluding Transkun is 'the most robust model.' Since Fig. 2b shows velocity has the highest cross-performer variance, the missing robustness table is load-bearing for the downstream claim, but this is a correctness/evidence concern, not constructional circularity. The paper's own admission of the omitted comparison is not a reduction of the result to its inputs. Overall: no significant circularity; score 2 reflects the minor self-citation/omitted-support context.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim rests on five domain assumptions, none of which are formal or machine-checked. There are no invented entities. The only data-derived parameters are the per-piece score-state anchor and model hyperparameters. The load-bearing assumptions are the reliability of the annotations, the fidelity of transcription velocities, and the absence of piece-identity leakage through the note-aligned codec.

free parameters (3)
  • per-piece score-state mu_score (valence, arousal) = not reported; centroid of training annotations per piece
    Eq. (1)-(2): every Delta-VA prediction adds this data-derived anchor; without a zero-deviation baseline, the model's R2=0.786 cannot be separated from the anchor's contribution.
  • Delta-VA hyperparameters = embedding dim 16; shared MLP widths; two output heads; RnC loss weighting
    Sections 5.2/5.3: chosen by hand; no ablations or sensitivity analysis reported.
  • segment length N=314 with cyclic padding = 314 notes
    Section 5: maximum annotated segment length; padding scheme may affect sequence models, but no analysis of its effect is given.
axioms (5)
  • domain assumption The four performance-codec features (beat period, velocity, timing, articulation) capture the expressive variation relevant to perceived emotion.
    Section 3.2: features adopted from [6]; Fig. 2 shows they vary across performers but does not prove they carry the emotion-relevant variation.
  • domain assumption The valence-arousal annotations of the first 8 bars are reliable, noiseless measures of perceived emotion.
    Section 3.1: dataset from [18]; no inter-annotator agreement or rating noise reported.
  • domain assumption Transkun transcription preserves velocity dynamics sufficiently for emotion prediction.
    Section 3.3: claim made; supporting comparison table omitted for space.
  • domain assumption The score-state centroid is a meaningful reference for an 'average' performance of a piece.
    Section 5.3: defines deviations from this centroid; the validity of this reference is not tested.
  • domain assumption Same-piece, different-performer training/test splits do not allow the model to exploit piece identity.
    Section 3.1: split by performer, so all 48 pieces occur in both train and test; codec is aligned to score notes, so piece structure is implicitly present.

pith-pipeline@v1.3.0-alltime-deepseek · 10463 in / 14656 out tokens · 166338 ms · 2026-08-03T01:25:02.117919+00:00 · methodology

0 comments
read the original abstract

Music is often used as a medium for communicating emotion, with performers shaping perceived affect through interpretation. This study addresses the challenge of identifying and predicting subtle changes in perceived emotion that are exclusively due to differences in performance. We focus on classical solo piano music, using a set of 6 commercial recordings of Bach's Well-Tempered Clavier Book I, annotated in terms of valence and arousal. By encoding the recordings through performance-specific features only, we isolate performance information from aspects of the composition itself, which tend to dominate the overall perceived emotional category. A preliminary analysis validates that these features vary meaningfully across performers. We then propose a relative regression framework, Delta-VA, to predict deviations in valence-arousal relative to an ``average'' performance, thereby focusing on the changes in emotion brought about by a specific way of playing a piece. In addition to the standard $R^2$ regression score, we introduce geometric evaluation metrics to assess the preservation of pairwise differences between performances. Results indicate high directional consistency with the ground truth, but also a compression in prediction magnitude, indicating that the model tends to underestimate expressive performance effects.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 3 linked inside Pith

  1. [1]

    Learning to Predict Performance- induced Emotion Differences in Classical Piano Music

    INTRODUCTION Music Emotion Recognition (MER) has been a continuing research topic in MIR, due to the central role of emotion and affect in music appreciation and the numerous applica- tion possibilities. The general approaches – training classi- fiers on top of hand-selected audio features, or using deep neural networks directly on the recordings (e.g., [...

    Pith/arXiv arXiv 2026

  2. [2]

    melodiousness

    RELA TED WORK Although performance-based features have not been widely used for emotion modeling, they have been exten- sively studied in research on computational performance modeling (e.g., [7, 8]). Similarly, expressive interpreta- tion across performers has been investigated, e.g., in the Mazurka Project [9]. Performance features are also com- monly e...

  3. [3]

    DA TA REPRESENTA TION As discussed above, when considering feature-based rep- resentations of music, we distinguish betweenscore fea- turesandperformance features, where the former relate to aspects of the piece, the composition itself (e.g., pitches, meter, mode, key), and the latter capture aspects shaped by the performer’s interpretation (e.g., precise...

  4. [4]

    To ensure that the performance-based features capture meaningful variation across performers rather than mainly encoding piece iden- tity, we conduct the following analyses

    ANALYSIS OF FEA TURE V ARIABILITY As illustrated in Figure 1, different performances of the same piece exhibit similar structural patterns, as they share the same underlying composition. To ensure that the performance-based features capture meaningful variation across performers rather than mainly encoding piece iden- tity, we conduct the following analys...

  5. [5]

    informa- tion

    LEARNING FROM PERFORMANCE We now turn to our main goal – predicting perceived emo- tion from performance details – and compare two formu- lations: (i) standard regression models that predict abso- lute valence–arousal values (baselines), and (ii) a relative regression formulation that predicts deviations induced by performance variations. This distinction...

  6. [6]

    give me a performance that is more positive and less aggressive than X

    ANALYSIS As mentioned in Section 5, all models – includingDelta- VA, are evaluated by MSE andR 2-score (Table 1). How- ever, simply assessing the performance by a highR2-score can be misleading, especially in our setting of predicting performance-induced deviations. While it indicates that the predictions are numerically close to the ground truth, it does...

  7. [7]

    CONCLUSION In this work, we investigated Music Emotion Recognition utilizing performance-related features alone. In particular, we proposed the Delta-V A framework to learn from perfor- mance features to predict deviations in perceived valence and arousal that are an effect of differences in expressive interpretation. Through variance analyses and represe...

  8. [8]

    101019375 (Whither Music?)

    ACKNOWLEDGMENTS This work is supported by the European Research Council (ERC) under the EU’s Horizon 2020 research & innovation programme, grant agreement No. 101019375 (Whither Music?)

    2020

  9. [9]

    Computational analysis and modeling of ex- pressive timing in chopin’s mazurkas

    Z. Shi, “Computational analysis and modeling of ex- pressive timing in chopin’s mazurkas.” inProceedings of the 22nd International Society for Music Informa- tion Retrieval (ISMIR), 2021

    2021

  10. [10]

    Rendering music performance with interpre- tation variations using conditional variational rnn,

    A. M. K. Y . T. Fujishima, K. Maezawa, and T. Ya- mamoto, “Rendering music performance with interpre- tation variations using conditional variational rnn,” in Proceedings of the 20th International Society for Mu- sic Information Retrieval (ISMIR), Delft, The Nether- lands, 2019

    2019

  11. [11]

    Regression-based music emotion prediction using triplet neural networks,

    K. W. Cheuk, Y .-J. Luo, B. B. T., G. Roig, and D. Her- remans, “Regression-based music emotion prediction using triplet neural networks,”International Joint Con- ference on Neural Networks, 2020

    2020

  12. [12]

    Stacked convolutional and recurrent neural networks for music emotion recogni- tion,

    M. Malík, S. Adavanne, K. Drossos, T. Virtanen, D. Ticha, and R. Jarina, “Stacked convolutional and recurrent neural networks for music emotion recogni- tion,” inProceedings of the 14th Sound & Music Com- puting Conference, 2017, p. 208–213

    2017

  13. [13]

    Bidirectional convolutional recurrent sparse network (BCRSN): An efficient model for music emotion recognition,

    Y . Dong, X. Yang, X. Zhao, and J. Li, “Bidirectional convolutional recurrent sparse network (BCRSN): An efficient model for music emotion recognition,”IEEE Transactions on Multimedia, vol. 21, pp. 3150–3163, 2019

    2019

  14. [14]

    Emotional expression in music performance: Between the performer’s inten- tion and the listener’s experience,

    A. Gabrielsson and P. N. Juslin, “Emotional expression in music performance: Between the performer’s inten- tion and the listener’s experience,”Psychology of Mu- sic, vol. 24, no. 1, pp. 68–91, 1996

    1996

  15. [15]

    Decoding emotions in expressive music performances: A multi-lab replication and extension study,

    J. Akkermans, R. Schapiro, D. Müllensiefen, K. Jakubowski, D. Shanahan, D. Baker, V . Busch, K. Lothwesen, P. Elvers, T. Fischinger, K. Schlemmer, and K. Frieler, “Decoding emotions in expressive music performances: A multi-lab replication and extension study,”Cognition and Emotion, vol. 33, no. 6, pp. 1099–1118, 2019

    2019

  16. [16]

    Computational Modeling of Expressive Music Performance with Linear and Non- Linear Basis Function Models,

    C. E. Cancino-Chacón, “Computational Modeling of Expressive Music Performance with Linear and Non- Linear Basis Function Models,”Johannes Kepler Uni- versity, Linz, 2018

    2018

  17. [17]

    Linear basis models for prediction and analysis of musical expression,

    M. Grachten and G. Widmer, “Linear basis models for prediction and analysis of musical expression,”Journal of New Music Research, vol. 41, no. 4, pp. 311–322, 2012

    2012

  18. [18]

    An evaluation of linear and non-linear models of expressive dynamics in classical piano and symphonic music,

    C. E. Cancino-Chacón, T. Gadermaier, G. Widmer, and M. Grachten, “An evaluation of linear and non-linear models of expressive dynamics in classical piano and symphonic music,”Machine Learning, vol. 106, no. 6, pp. 887–909, 2017

    2017

  19. [19]

    Measurement and Reproduc- tion Accuracy of Computer-controlled Grand Pianos,

    W. Goebl and R. Bresin, “Measurement and Reproduc- tion Accuracy of Computer-controlled Grand Pianos,” The Journal of the Acoustical Society of America, vol. 114, no. 4, pp. 2273–2283, 10 2003

    2003

  20. [20]

    A Statistical View on the Expressive Timing of Piano Rolled Chords,

    M. Fu, G. Xia, R. B. Dannenberg, and L. A. Wasser- man, “A Statistical View on the Expressive Timing of Piano Rolled Chords,” inProceedings of the 16th Inter- national Society for Music Information Retrieval Con- ference (ISMIR), Málaga, Spain, 2015

    2015

  21. [21]

    Vir- tuosonet: A hierarchical rnn-based system for model- ing expressive piano performance

    D. Jeong, T. Kwon, Y . Kim, K. Lee, and J. Nam, “Vir- tuosonet: A hierarchical rnn-based system for model- ing expressive piano performance.” inProceedings of the 20th International Society for Music Information Retrieval (ISMIR), Delft, The Netherlands, 2019

    2019

  22. [22]

    Scoreperformer: Expressive piano performance rendering with fine-grained con- trol

    I. Borovik and V . Viro, “Scoreperformer: Expressive piano performance rendering with fine-grained con- trol.” inProceedings of the 24th International Society for Music Information Retrieval (ISMIR), Milan, Italy, 2023

    2023

  23. [23]

    DExter: Learning and controlling performance expression with diffusion models,

    H. Zhang, S. Chowdhury, C. E. Cancino-Chacón, J. Liang, S. Dixon, and G. Widmer, “DExter: Learning and controlling performance expression with diffusion models,”Applied Sciences, 2024

    2024

  24. [24]

    The performance of music,

    A. Gabrielsson, “The performance of music,” inThe psychology of music. Elsevier, 1999, pp. 501–602

    1999

  25. [25]

    Acoustically Expressing Affect,

    A. Battcock and M. Schutz, “Acoustically Expressing Affect,”Music Perception, vol. 37, no. 1, pp. 66–91, 09 2019

    2019

  26. [26]

    Individualized interpretation: Exploring struc- tural and interpretive effects on evaluations of emo- tional content in bach’s well tempered clavier,

    ——, “Individualized interpretation: Exploring struc- tural and interpretive effects on evaluations of emo- tional content in bach’s well tempered clavier,”Journal of New Music Research, vol. 50, pp. 447 – 468, 2021

    2021

  27. [27]

    Beyond the notes: Clarifying the role of expressivity in conveying musical emotion,

    C. J. Anderson, J. Ling, and M. Schutz, “Beyond the notes: Clarifying the role of expressivity in conveying musical emotion,”Quarterly Journal of Experimental Psychology, 2025

    2025

  28. [28]

    On perceived emotion in expressive piano performance: Further experimental evidence for the relevance of mid-level perceptual fea- tures,

    S. Chowdhury and G. Widmer, “On perceived emotion in expressive piano performance: Further experimental evidence for the relevance of mid-level perceptual fea- tures,” inProceedings of the 22nd International Society for Music Information Retrieval (ISMIR), 2021

    2021

  29. [29]

    andmpteval[30]. Rather than absolute accuracy, we prioritize correlation-based metrics, as relative loudness variation is more important for describing expressive dy- namics.mir_eval, widely used for transcription evalua- tion, reports metrics for velocity-only and note+velocity, with the latter computed only for correctly matched notes based on pitch and...

  30. [30]

    Towards Musically Informed Evaluation of Pi- ano Transcription Models,

    P. Hu, L. S. Marták, C. Cancino-Chacón, and G. Wid- mer, “Towards Musically Informed Evaluation of Pi- ano Transcription Models,” inProceedings of the 25th International Society for Music Information Retrieval Conference (ISMIR), San Francisco, USA, 2024

    2024

  31. [31]

    Melody Lead in Piano Performance: Ex- pressive Device or Artifact?

    W. Goebl, “Melody Lead in Piano Performance: Ex- pressive Device or Artifact?”The Journal of the Acoustical Society of America, vol. 110, no. 1, pp. 563– 72, 2001

    2001

  32. [32]

    Parti- tura: A Python Package for Symbolic Music Process- ing,

    C. E. Cancino-Chacón, S. D. Peter, E. Karystinaios, F. Foscarin, M. Grachten, and G. Widmer, “Parti- tura: A Python Package for Symbolic Music Process- ing,” inProceedings of the Music Encoding Conference (MEC), Halifax, Canada, 2022

    2022

  33. [33]

    A Data-Driven Analysis of Ro- bust Automatic Piano Transcription,

    D. Edwards, S. Dixon, E. Benetos, A. Maezawa, and Y . Kusaka, “A Data-Driven Analysis of Ro- bust Automatic Piano Transcription,”arXiv preprint arXiv:2402.01424, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Sound and mu- sic biases in deep music transcription models: a sys- tematic analysis,

    L. Martak, P. Hu, and G. Widmer, “Sound and mu- sic biases in deep music transcription models: a sys- tematic analysis,”EURASIP Journal on Audio, Speech, and Music Processing, no. 1, 2026

    2026

  35. [35]

    ASAP: a dataset of aligned scores and per- formances for piano transcription,

    F. Foscarin, A. McLeod, P. Rigaux, F. Jacquemard, and M. Sakai, “ASAP: a dataset of aligned scores and per- formances for piano transcription,” inProceedings of the 21st International Society for Music Information Retrieval Conference (ISMIR), 2020, pp. 534–541

    2020

  36. [36]

    High-resolution Piano Transcription with Pedals by Regressing Onsets and Offsets Times,

    Q. Kong, B. Li, X. Song, Y . Wan, , and Y . Wang, “High-resolution Piano Transcription with Pedals by Regressing Onsets and Offsets Times,”arXiv preprint arXiv:2010.01815, 2020

    Pith/arXiv arXiv 2010

  37. [37]

    Automatic Piano Transcrip- tion with Hierarchical Frequency-Time Transformer,

    K. Toyama, T. Akama, Y . Ikemiya, Y . Takida, W.-H. Liao, and Y . Mitsufuji, “Automatic Piano Transcrip- tion with Hierarchical Frequency-Time Transformer,” inProceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR), Mi- lan, Italy, 2023

    2023

  38. [38]

    Scoring Time Intervals Us- ing Non-Hierarchical Transformer for Automatic Pi- ano Transcription,

    Y . Yan and Z. Duan, “Scoring Time Intervals Us- ing Non-Hierarchical Transformer for Automatic Pi- ano Transcription,” inProceedings of the 25th Inter- national Society for Music Information Retrieval Con- ference (ISMIR), San Francisco, USA, 2024

    2024

  39. [39]

    MIR_EV AL: A Transparent Implementation of Common MIR Met- rics,

    C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Nieto, D. Liang, and D. P. W. Ellis, “MIR_EV AL: A Transparent Implementation of Common MIR Met- rics,” inInternational Society for Music Information Retrieval Conference, Taipei, Taiwan, 2014

    2014

  40. [41]

    Representational similarity analysis - connecting the branches of systems neuroscience,

    N. Kriegeskorte, M. Mur, and P. A. Bandettini, “Representational similarity analysis - connecting the branches of systems neuroscience,”Frontiers in Sys- tems Neuroscience, vol. 2, 2008

    2008

  41. [42]

    Rank-n-contrast: learning continuous representations for regression,

    K. Zha, P. Cao, J. Son, Y . Yang, and D. Katabi, “Rank-n-contrast: learning continuous representations for regression,” inProceedings of the 37th Interna- tional Conference on Neural Information Processing Systems, Red Hook, NY , USA, 2023

    2023

  42. [127]

    Although the relationship between loudness and MIDI velocity is not strictly linear [19], normal- ized MIDI velocity is used as a proxy for dynamics. •Timing. This parameter captures temporal devia- tions from the nominal score timing, including ex- pressive phenomena such as chord spread [20] and melody lead [21]. Positive values indicate antici- pation,...