Pith. sign in

REVIEW 3 major objections 4 minor 68 references

Improving BERT for Symbolic Music Understanding Using Token Denoising and Pianoroll Prediction

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that bounded token denoising and pianoroll prediction pre-training objectives improve symbolic music understanding in a BERT-style encoder, outperforming MidiBERT on almost all of twelve downstream tasks.

desk verdict A useful 12-task benchmark and two reasonable pre-training ideas, but the effectiveness claim outruns the statistics without error bars or a pianoroll-only ablation. read the letter →

arxiv 2507.04776 v1 pith:UWDAKSOI submitted 2025-07-07 cs.SD cs.LGcs.MMeess.AS

Jun-You Wang , Li Su This is my paper · ORCID
classification cs.SDcs.LGcs.MMeess.AS
keywords symbolicmusicunderstandingpre-trainedmodelsBERTtokendenoisingpianorollpredictionself-supervisedlearninginformationretrievaldownstreamevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that two self-supervised pre-training objectives—bounded token denoising and pianoroll prediction—make a BERT-style model (M2BERT, built on a 12-layer ModernBERT) better at understanding symbolic music than the standard masked-language-model approach used by MidiBERT. The authors argue that plain MLM treats notes as isolated tokens and never teaches the model the musical relationships between them, such as pitch intervals and metrical structure. Over a new benchmark of 12 downstream classification tasks, the best M2BERT configuration outperforms MidiBERT on almost all tasks, with the largest gains on downbeat prediction, chord root estimation, local key estimation, and melody extraction. If correct, this points to pre-training objectives that encode musical geometry rather than larger architectures alone as the key to stronger symbolic music models.

What carries the argument

The two pre-training objectives. Token denoising corrupts a note token's attributes by bounded random noise—onset position within ±4, pitch within ±12 semitones, duration within ±12 units—and trains the encoder with cross-entropy loss to recover the original attributes, with the bound serving as the mechanism that forces the model to reason about near-miss values and thereby learn interval structure. Pianoroll prediction adds an L2 loss that requires each note embedding to predict a 16-by-86 pianoroll and a 16-by-12 chromagram for its bar (bar-level) and for its own onset time (local), directly injecting the simultaneous-pitch layout of the music into the learned representation. Both objectives share the same note-level contextualized encoder and prediction machinery, so the pre-training signal is a sum of token reconstruction loss and pianoroll prediction loss.

What would settle it

Pre-train the same model and data but compute the pianoroll targets from randomly permuted downbeat timings; if downstream gains on chord and melody tasks survive, the pianoroll objective is not doing the work the paper claims.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing masked language modeling with two music-aware objectives makes a BERT-style encoder better at symbolic music understanding. In the token denoising objective, roughly 30% of note tokens have their onset position, pitch, and duration attributes perturbed within bounded ranges (position by ±4, pitch and duration by ±12), and the model learns to reconstruct the original attribute values, which teaches the relative distances between token values that carry musical meaning such as pitch intervals. In the pianoroll prediction objective, each note's embedding is additionally trained with an L2 loss to reproduce the pianoroll and chromagram of its enclosing bar and of its own onset position, injecting explicit information about which pitches sound together and when. Combined into a 12-layer ModernBERT backbone and pre-trained on up to 350 million notes, the resulting M2BERT model outperforms MidiBERT on almost all twelve downstream tasks, with the largest gains on downbeat prediction, chord root estimation, local key estimation, and melody extraction.

Load-bearing premise

The method assumes that downbeat timings derived from MIDI tick information are correct enough that the pianoroll and chroma targets align with the actual music, because those targets are computed from bars defined by the downbeats.

Editorial extensions

If this is right

  • Bounded denoising (RC4,12,12) outperforms both standard MLM and unbounded random replacement (RC∞), showing that limiting the corruption range is itself a source of the improvement.
  • Adding the pianoroll head yields the largest gains on tasks that depend on pitch, interval, and harmonic structure, such as chord root, local key, melody, texture, and downbeat prediction, while leaving beat-note prediction and velocity estimation roughly unchanged.
  • Scaling the pre-training corpus from 4.89 million to 350.32 million notes improves nearly every downstream task, and increasing the input context length from 512 to 2048 tokens helps most on chord and key estimation.
  • The best configuration reaches or beats previously reported task-specific state-of-the-art results on several tasks without task-specific architectures, data augmentation, or multi-task learning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The bounded-denoisng principle should transfer to other structured-sequence domains whose tokens live on an ordered metric, where corrupting within a limited radius and denoising would teach distance structure analogous to musical intervals.
  • Since the pianoroll target is computed from downbeats, coupling the model with a learned beat tracker, or treating the downbeat as a predicted variable, would be a natural robustness extension and a direct stress test of the objective.
  • The SMC benchmark's focus on classification tasks may under-represent what the objectives do for generation or score-to-performance mapping; those settings could show different trade-offs.
  • The concentration of gains in pitch-interval-sensitive tasks yields a testable prediction: corrupting only duration and onset while keeping pitch intact should not reproduce the chord and melody improvements.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes M2BERT, a ModernBERT-based symbolic-music model, together with two pre-training objectives: bounded token denoising (RC4,12,12), which corrupts note attributes with limited noise and trains the model to reconstruct them, and pianoroll prediction, which trains the model to predict bar-level and local pianoroll/chroma targets. The authors also introduce the SMC benchmark, comprising 12 symbolic-music classification tasks with a unified fine-tuning protocol, and report experiments comparing MidiBERT, several M2BERT variants, two pre-training corpus sizes, and different maximum sequence lengths. The main empirical claim is that the proposed objectives improve downstream symbolic-music understanding, with the best configuration reaching or exceeding previous SOTA numbers on several tasks.

Significance. If the empirical claim is robust, the paper makes a useful contribution: it offers two principled, music-aware pre-training objectives, a reproducible 12-task benchmark, and a careful set of ablations that separate bounded from unbounded token corruption. The benchmark and the public code/data links are concrete assets for the community. The bounded-denoising idea is a reasonable middle ground between masking and full random replacement, and the sequence-length ablation gives useful practical guidance. However, the central claim is currently supported by single-run numbers without uncertainty estimates, and the pianoroll objective is never evaluated by itself, so the evidence is not yet at the level the abstract claims.

major comments (3)
  1. [§6, Table 1] The paper's headline claim, stated in the abstract as "Results confirm the effectiveness of the proposed pre-training objectives," rests on comparisons in Table 1 that are reported without any uncertainty quantification. For the comparison that isolates the pianoroll objective, M2BERT RC4,12,12+Pianoroll versus M2BERT RC4,12,12, the differences are mostly a few thousandths and are negative on four tasks (DbP −0.011, ME −0.001, VE −0.002, MNID −0.001) and zero on BP. Even the combined-objective comparison against the same-architecture MLM baseline shows gains of at most 0.030 on most tasks, and no standard deviations, confidence intervals, number of seeds, or significance tests are provided anywhere in Section 6. At this scale the observed differences could plausibly be seed noise. Please report multiple fine-tuning runs (at least 3–5 seeds) with means and standard deviations, or a paired significance test, and adjust the strength of the conclusions to what the statistics support.
  2. [§3.3.2, Table 1] The pianoroll prediction objective is never trained alone. The comparison RC4,12,12+Pianoroll versus RC4,12,12 isolates only the marginal effect of adding pianoroll prediction on top of token denoising; it cannot establish the individual contribution of the pianoroll objective, which the paper presents as one of its two main contributions. A Pianoroll-only condition, such as MLM+Pianoroll or token-denoising-free pianoroll prediction, is needed to determine whether the objective is beneficial by itself or only in interaction with token denoising. Without that condition, the claim that the proposed objectives improve performance is not fully supported.
  3. [§3.1 and §3.3.2] Both the tokenization and the pianoroll targets rely critically on downbeat timings DB. Section 3.1 says that for MIDI data, "an unreliable yet doable way" is to retrieve tick information, and the pre-training corpus includes performance MIDI such as Pianist8 and ASAP, where such information may be noisy or absent. If the downbeat estimates are wrong, the bar boundaries, the rescaling to four crotchet beats, and the pianoroll/chroma targets all become misaligned with the actual musical content. This directly threatens the mechanism of the pianoroll objective. Please either quantify the reliability of DB estimation on the pre-training corpus, restrict the corpus to reliable downbeat annotations, or demonstrate robustness to small downbeat perturbations, before concluding that the pianoroll objective is effective.
minor comments (4)
  1. [§6, SOTA discussion] The sentence listing tasks where the full model "outperforms the SOTAs" is internally inconsistent: it enumerates seven tasks (DbP, CR, LK, ME, OTC, PS, VF) but says "six of the ten tasks," and LK is not actually beaten in Table 1 (0.811 versus SOTA 0.829). Please correct the enumeration and the count.
  2. [§5] Section 5 says the model "randomly corrupts 30% of the note tokens," while Section 3.3.1 describes MLM as using 15% of tokens in practice. Please state explicitly whether the 30% ratio applies to all compared conditions, including the MidiBERT MLM baseline and the M2BERT MLM baseline; if it does not, the M2BERT-versus-MidiBERT comparison in Section 6 partly conflates architecture with corruption ratio.
  3. [§3.3.1, Eqs. (2)–(4)] The notation rand(clip(posn − rpos), clip(posn + rpos)) is slightly informal: it should be made explicit that clip returns an integer endpoint in the valid token range and that the draw is over integers in the closed interval [x, y]. The limiting case rpos = rpit = rdur = ∞ also deserves a short formal definition rather than an appeal to the reader's intuition.
  4. [§3.3.2, Eq. (5) and following text] There is a missing space in "pianoroll predictionobjective," and the line break in "PRn and CMn" makes the expression harder to read. These are purely typographical issues.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: self-supervised objectives are evaluated on external downstream tasks with no fitted parameter that defines the result.

full rationale

The paper's central claim is that two pre-training objectives (bounded token denoising and pianoroll prediction) improve downstream symbolic-music classification. This is an empirical, self-contained claim. The pre-training targets are derived from the input tokens themselves (denoised original tokens; pianoroll/chroma computed from the same bar), which is standard self-supervision rather than circular reasoning: the model is asked to predict deterministic functions of its own input, and the evaluation is performed on held-out downstream tasks with public datasets and fixed metrics. The corruption ranges (rpos=4, rpit=12, rdur=12) are hand-chosen hyperparameters, not fitted to any downstream label, so the gains on DbP, CR, ME, OTC, and ER are not forced by construction. The benchmark (SMC) is newly assembled, but it consists of existing datasets and tasks (Tagtraum, Pianist8, EMOPIA, PM2S, AugmentedNet, POP909, BPS-motif, TNUA, etc.) with external SOTA references, so the evaluation does not reduce to the paper's own definitions. Some cited works share authors (e.g., BPS-motif [18], OTC [51], VF [53]), but these citations provide datasets or baseline methods, not the load-bearing justification for the proposed objectives; removing them would not make the objectives self-validating. The paper even reports failures (VF is worse than MidiBERT) and notes nonuniform gains, which is inconsistent with a result that is circularly guaranteed. No equation equates a downstream metric with a pre-training loss, and no fitted parameter is renamed as a prediction. The main weakness, lack of significance tests and error bars, is a statistical-evidence concern, not a circularity concern.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several hand-chosen hyperparameters (corruption ranges and ratio, sequence length) and on domain assumptions about the reliability of MIDI downbeat information and the validity of the 4/4 rescaling. No new theoretical entities or unverified physical/factual objects are introduced.

free parameters (5)
  • rpos (corruption range for onset position) = 4
    Hand-chosen hyperparameter in Eq. (2). Not derived from downstream performance; selected by the authors.
  • rpit (corruption range for pitch) = 12
    Hand-chosen hyperparameter in Eq. (3). The authors argue it guides pitch-interval learning, but the value is not tuned against downstream tasks.
  • rdur (corruption range for duration) = 12
    Hand-chosen hyperparameter in Eq. (4).
  • token corruption ratio = 0.30
    Section 5: 'randomly corrupt 30% of the note tokens for pre-training'. Hand-selected.
  • maximum sequence length = 1024
    Selected after ablation (Table 2) as a trade-off between performance and complexity; not automatically tuned.
assumptions (4)
  • domain assumption MIDI tick information reliably provides downbeat timings (DB) for the pre-training data.
    Section 3.1: 'an unreliable yet doable way to obtain the raw DB information is to retrieve the tick information'. The tokenization and pianoroll targets depend on this.
  • domain assumption All bars can be rescaled to four crotchet beats (4/4) without losing information essential to downstream tasks.
    Section 3.1: 'we follow MidiBERT by assuming that the duration of a bar is four crotchet beats... For bars that have a different duration... we rescale them'.
  • domain assumption The pianoroll and chroma targets are computed from the same note sequence and their L2 prediction is a useful learning signal.
    Section 3.3.2 defines the targets from the input tokens; the paper hypothesizes that predicting them teaches pitch intervals and metrical knowledge.
  • domain assumption The 12 downstream tasks and datasets are representative of symbolic music understanding.
    Section 4 argues for diversity, comparability, and reproducibility, but the selection is the authors' choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving BERT for Symbolic Music Understanding Using Token Denoising and Pianoroll Prediction." pith.science (2026). https://pith.science/paper/UWDAKSOI

@misc{pith2026250704776,
  author       = {Pith},
  title        = {Pith review of: Improving BERT for Symbolic Music Understanding Using Token Denoising and Pianoroll Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWDAKSOI}},
  note         = {Machine review of arXiv:2507.04776}
}
read the original abstract

We propose a pre-trained BERT-like model for symbolic music understanding that achieves competitive performance across a wide range of downstream tasks. To achieve this target, we design two novel pre-training objectives, namely token correction and pianoroll prediction. First, we sample a portion of note tokens and corrupt them with a limited amount of noise, and then train the model to denoise the corrupted tokens; second, we also train the model to predict bar-level and local pianoroll-derived representations from the corrupted note tokens. We argue that these objectives guide the model to better learn specific musical knowledge such as pitch intervals. For evaluation, we propose a benchmark that incorporates 12 downstream tasks ranging from chord estimation to symbolic genre classification. Results confirm the effectiveness of the proposed pre-training objectives on downstream tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages

  1. [1]

    Improving BERT for Symbolic Music Understanding Using Token Denoising and Pianoroll Prediction

    INTRODUCTION In recent years, music information retrieval (MIR) research in the symbolic music domain has undergone a paradigm shift from the development of task-specific models to the adoption of the pre-training/fine-tuning paradigm, where a model is first pre-trained on a large-scale dataset with self- supervised learning (SSL) objectives and then fine...

    work page Pith review arXiv 2025

  2. [2]

    Each note contains at least three attributes: onset timing, note duration, and note pitch

    RELATED WORK Symbolic music data is an abstract representation of mu- sic [17], which represents a musical piece with a sequence of notes. Each note contains at least three attributes: onset timing, note duration, and note pitch. Additional attributes such as tempo, time signature, staff information, etc., may also present. Symbolic music understanding ai...

    work page

  3. [3]

    the pitch difference between E4 and C4 is necessarily a major third

    PROPOSED METHOD Figure 1 gives an overview of the proposed method. The input symbolic music representation is a music piece X composed of N notes, i.e., X := {xn}N n=1, and xn := (o n, pn, dn) represents the n-th note with its onset, pitch, and duration in X, and a list of down- beat timings DB composed of M downbeat timings, i.e., DB := {dbm}M m=1. For M...

    work page

  4. [4]

    We refer to the combination of all these tasks and datasets as the SMC benchmark, which stands for Symbolic Music Classifica- tion benchmark

    THE SMC BENCHMARK We conduct an evaluation on 12 different downstream tasks for symbolic music classification. We refer to the combination of all these tasks and datasets as the SMC benchmark, which stands for Symbolic Music Classifica- tion benchmark. The benchmark is available at https: //zenodo.org/records/15681035. Similar to [1, 3, 4, 11], we focus o...

    work page arXiv

  5. [5]

    no label

    determine the global tempo T by assuming the length of one beat as the median value of all notes’ duration; 2) if T /∈ [40, 200], then T is multiplied or divided by 2 to make it lie in the range; 3) quantize the note sequence to 16th notes by assuming a 4/4 time signature and constant tempo regardless of the actual performance content. 3 Downbeat note pre...

    work page

  6. [6]

    The cor- ruption hyperparameters r pos, r pit, and r dur are set to 4, 12, and 12, respectively

    EXPERIMENT SETUP We utilize the StableAdamW optimizer [54] and randomly corrupt 30% of the note tokens for pre-training. The cor- ruption hyperparameters r pos, r pit, and r dur are set to 4, 12, and 12, respectively. All other pre-training hyperpa- rameters remain the same as MidiBERT, namely: learning rate of 2 × 10−5, weight decay of 0.01, batch size o...

    work page

  7. [7]

    For reference, we show the state-of-the-art (SOTA) performance of the tasks in the last row, if available

    RESULTS Table 1 presents the experiment results over all the settings on the 12 downstream tasks. For reference, we show the state-of-the-art (SOTA) performance of the tasks in the last row, if available. First, we observe that all the settings of the proposed M2BERT model outperforms MidiBERT for almost all the downstream tasks; this demonstrates the ad-...

    work page

  8. [8]

    CONCLUSION With a systematic evaluation on a benchmark incorporat- ing 12 downstream tasks, we have demonstrated the ef- fectiveness of using token denoising and pianoroll predic- tion to enhance the pre-training of a BERT-like model for symbolic music understanding. This result underscores the important insight that symbolic music pre-training should foc...

    work page

Show all 68 references
  1. [9]

    ETHICS STATEMENT In this work, all the datasets used in model training and evaluation are publicly available and can be down- loaded without submitting any data access application form. While this strict policy improves the reproducibility, it also makes the evaluation biased ...

  2. [10]

    ACKNOWLEDGMENTS This work is supported in part by National Science and Technology Council under Grant NSTC 113-2221-E-001- 013, the Academia Sinica Grand Challenge (GCS) Pro- gram under Grant AS–GCS–112–M07, and the Postdoc- toral Scholar Program of Academia Sinica under Grant...

  3. [11]

    MidiBERT-Piano: Large-scale pre-training for symbolic music understanding,

    Y . Chou, I. Chen, C. Chang, J. Ching, and Y . Yang, “MidiBERT-Piano: Large-scale pre-training for symbolic music understanding,” CoRR, vol. abs/2107.05223, 2021

    2021 arXiv

  4. [12]

    PianoBART: Symbolic piano mu- sic generation and understanding with large-scale pre- training,

    X. Liang et al. , “PianoBART: Symbolic piano mu- sic generation and understanding with large-scale pre- training,” in IEEE International Conference on Multi- media and Expo, ICME, 2024, pp. 1–6

    2024

  5. [13]

    Adversarial-midibert: Symbolic music un- derstanding model based on unbias pre-training and mask fine-tuning,

    Z. Zhao, “Adversarial-midibert: Symbolic music un- derstanding model based on unbias pre-training and mask fine-tuning,” CoRR, vol. abs/2407.08306, 2024

    2024 arXiv

  6. [14]

    More than simply masking: Exploring pre-training strategies for symbolic music understanding,

    Z. Shen, L. Yang, Z. Yang, and H. Lin, “More than simply masking: Exploring pre-training strategies for symbolic music understanding,” in Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, 2023, p. 540–544

    2023

  7. [15]

    MusicBERT: Symbolic music understanding with large-scale pre-training,

    M. Zeng, X. Tan, R. Wang, Z. Ju, T. Qin, and T. Liu, “MusicBERT: Symbolic music understanding with large-scale pre-training,” in Findings of the Asso- ciation for Computational Linguistics: ACL/IJCNLP , 2021, pp. 791–800

    2021

  8. [16]

    MuseBERT: Pre-training mu- sic representation for music understanding and control- lable generation,

    Z. Wang and G. Xia, “MuseBERT: Pre-training mu- sic representation for music understanding and control- lable generation,” in Proceedings of the 22nd Interna- tional Society for Music Information Retrieval Confer- ence, ISMIR 2021, 2021, pp. 722–729

    2021

  9. [17]

    RNBert: Fine-tuning a masked language model for roman numeral analysis,

    M. Sailor, “RNBert: Fine-tuning a masked language model for roman numeral analysis,” in Proceedings of the 25th International Society for Music Information Retrieval Conference, ISMIR, 2024, pp. 814–821

    2024

  10. [18]

    GPT-4 technical report,

    OpenAI, “GPT-4 technical report,” CoRR, vol. abs/2303.08774, 2023

    2023 arXiv

  11. [19]

    Llama: Open and efficient foun- dation language models,

    H. Touvron et al., “Llama: Open and efficient foun- dation language models,” CoRR, vol. abs/2302.13971, 2023

    2023 arXiv

  12. [20]

    Open LLM leaderboard,

    E. Beeching et al. , “Open LLM leaderboard,” https://huggingface.co/spaces/HuggingFaceH4/open_ llm_leaderboard, 2023

    2023

  13. [21]

    BERT: pre-training of deep bidirectional transform- ers for language understanding,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional transform- ers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Hu- man Language Technolog...

    2019

  14. [22]

    Melodic matching techniques for large music databases,

    A. L. Uitdenbogerd and J. Zobel, “Melodic matching techniques for large music databases,” in Proceedings of the 7th ACM International Conference on Multime- dia, 1999, pp. 57–66

    1999

  15. [23]

    A convolutional approach to melody line identification in symbolic scores,

    F. Simonetta, C. E. C. Chacón, S. Ntalampiras, and G. Widmer, “A convolutional approach to melody line identification in symbolic scores,” in Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR), 2019, pp. 924–931

    2019

  16. [24]

    A deep learning method for melody extraction from a polyphonic symbolic music representation,

    K. Kosta, W. T. Lu, G. Medeot, and P. Chanquion, “A deep learning method for melody extraction from a polyphonic symbolic music representation,” in Pro- ceedings of the 23rd International Society for Music Information Retrieval Conference (ISMIR) , 2022, pp. 757–763

    2022

  17. [25]

    Augment- edNet: A roman numeral analysis network with syn- thetic training examples and additional tonal tasks,

    N. N. López, M. Gotham, and I. Fujinaga, “Augment- edNet: A roman numeral analysis network with syn- thetic training examples and additional tonal tasks,” in Proceedings of the 22nd International Society for Music Information Retrieval Conference, ISMIR, 2021, pp. 404–411

    2021

  18. [26]

    Roman numeral anal- ysis with graph neural networks: Onset-wise predic- tions from note-wise features,

    E. Karystinaios and G. Widmer, “Roman numeral anal- ysis with graph neural networks: Onset-wise predic- tions from note-wise features,” in Proceedings of the 24th International Society for Music Information Re- trieval Conference, ISMIR, 2023, pp. 597–604

    2023

  19. [27]

    Natural language processing methods for symbolic music gen- eration and information retrieval: a survey,

    D. Le, L. Bigo, M. Keller, and D. Herremans, “Natural language processing methods for symbolic music gen- eration and information retrieval: a survey,”CoRR, vol. abs/2402.17467, 2024

    2024 arXiv

  20. [28]

    Bps-motif: A dataset for repeated pattern discovery of polyphonic symbolic music,

    Y . Hsiao, T. Hung, T. Chen, and L. Su, “Bps-motif: A dataset for repeated pattern discovery of polyphonic symbolic music,” in Proceedings of the 24th Interna- tional Society for Music Information Retrieval Confer- ence (ISMIR), 2023, pp. 281–288

    2023

  21. [29]

    2013:Discovery of Repeated Themes & Sections,

    T. Collins, “2013:Discovery of Repeated Themes & Sections,” https://www.music-ir.org/mirex/wiki/2013: Discovery_of_Repeated_Themes_%26_Sections, 2013

    2013

  22. [30]

    Algo- rithms for discovering repeated patterns in multidimen- sional representations of polyphonic music,

    D. Meredith, K. Lemström, and G. A. Wiggins, “Algo- rithms for discovering repeated patterns in multidimen- sional representations of polyphonic music,” Journal of New Music Research , vol. 31, no. 4, pp. 321–345, 2002

    2002

  23. [31]

    Functional harmony recogni- tion of symbolic music data with multi-task recurrent neural networks,

    T.-P. Chen and L. Su, “Functional harmony recogni- tion of symbolic music data with multi-task recurrent neural networks,” in Proceedings of the 19th Interna- tional Society for Music Information Retrieval Confer- ence (ISMIR), 2018, pp. 90–97

    2018

  24. [32]

    S3: A symbolic music dataset for computational music analysis of symphonies,

    Z.-S. Lin et al. , “S3: A symbolic music dataset for computational music analysis of symphonies,” in Ex- tended Abstracts for the Late-Breaking Demo Session of the 25th International Society for Music Information Retrieval Conference (ISMIR), 2024

    2024

  25. [33]

    Improving genre annotations for the million song dataset,

    H. Schreiber, “Improving genre annotations for the million song dataset,” inProceedings of the 16th Inter- national Society for Music Information Retrieval Con- ference, ISMIR, 2015, pp. 241–247

    2015

  26. [34]

    Symbolic music structure analysis with graph repre- sentations and changepoint detection methods,

    C. Hernandez-Olivan, S. R. Llamas, and J. R. Beltrán, “Symbolic music structure analysis with graph repre- sentations and changepoint detection methods,”CoRR, vol. abs/2303.13881, 2023

    2023 arXiv

  27. [35]

    EMOPIA: A multi-modal pop piano dataset for emotion recognition and emotion-based music gen- eration,

    H. Hung, J. Ching, S. Doh, N. Kim, J. Nam, and Y . Yang, “EMOPIA: A multi-modal pop piano dataset for emotion recognition and emotion-based music gen- eration,” in Proceedings of the 22nd International So- ciety for Music Information Retrieval Conference, IS- MIR, 2021, pp. 318–325

    2021

  28. [36]

    Multimodal multifaceted music emotion recognition based on self-attentive fusion of psychology-inspired symbolic and acoustic features,

    J. Zhao and K. Yoshii, “Multimodal multifaceted music emotion recognition based on self-attentive fusion of psychology-inspired symbolic and acoustic features,” in Asia Pacific Signal and Information Processing As- sociation Annual Summit and Conference, APSIPA ASC, 2023, pp. ...

    2023

  29. [37]

    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,” in Proceedings of the 21th International Society for Music Information Retrieval Conference, ISMIR, 2020, pp. 534–541

    2020

  30. [38]

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

    H. Zhang, J. Tang, S. R. Rafee, S. Dixon, G. Fazekas, and G. A. Wiggins, “ATEPP: A dataset of automat- ically transcribed expressive piano performance,” in Proceedings of the 23rd International Society for Mu- sic Information Retrieval Conference, ISMIR , 2022, pp. 446–453

    2022

  31. [39]

    A-MAPS: Augmented MAPS dataset with rhythm and key annotations,

    A. Ycart and E. Benetos, “A-MAPS: Augmented MAPS dataset with rhythm and key annotations,” in 19th International Society for Music Information Re- trieval Conference, ISMIR Late Breaking and Demo Papers, 2018

    2018

  32. [40]

    Can LLMs

    Z. Zhou et al. , “Can LLMs "reason" in music? an evaluation of llms’ capability of music understanding and generation,” in Proceedings of the 25th Interna- tional Society for Music Information Retrieval Confer- ence (ISMIR), 2024

    2024

  33. [41]

    ChatMusician: Understanding and generating music intrinsically with LLM,

    R. Yuan et al. , “ChatMusician: Understanding and generating music intrinsically with LLM,” in Findings of the Association for Computational Linguistics, ACL, 2024, pp. 6252–6271

    2024

  34. [42]

    On the opportunities and risks of foundation models,

    R. Bommasani et al., “On the opportunities and risks of foundation models,”CoRR, vol. abs/2108.07258, 2021

    2021 arXiv

  35. [43]

    A foundation model for music informatics,

    M. Won, Y . Hung, and D. Le, “A foundation model for music informatics,” in IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP , 2024, pp. 1226–1230

    2024

  36. [44]

    Learning transferable visual models from natural language supervision,

    A. Radford et al., “Learning transferable visual models from natural language supervision,” in Proceedings of the 38th International Conference on Machine Learn- ing, ICML, vol. 139, 2021, pp. 8748–8763

    2021

  37. [45]

    SUPERB: speech processing universal performance benchmark,

    S. Yang et al., “SUPERB: speech processing universal performance benchmark,” in Interspeech 2021, 22nd Annual Conference of the International Speech Com- munication Association, 2021, pp. 1194–1198

    2021

  38. [46]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  39. [47]

    Dynamic-SUPERB Phase-2: A collaboratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,

    C. Huang et al. , “Dynamic-SUPERB Phase-2: A collaboratively expanding benchmark for measuring the capabilities of spoken language models with 180 tasks,” in The Thirteenth International Conference on Learning Representations, ICLR 2025, 2025

    2025

  40. [48]

    MERT: acoustic music understanding model with large-scale self-supervised training,

    Y . Li et al. , “MERT: acoustic music understanding model with large-scale self-supervised training,” inThe Twelfth International Conference on Learning Repre- sentations, ICLR 2024, 2024

    2024

  41. [49]

    Music foundation model as generic booster for music downstream tasks,

    W. Liao et al. , “Music foundation model as generic booster for music downstream tasks,” CoRR, vol. abs/2411.01135, 2024

    2024 arXiv

  42. [50]

    Codified au- dio language modeling learns useful representations for music information retrieval,

    R. Castellon, C. Donahue, and P. Liang, “Codified au- dio language modeling learns useful representations for music information retrieval,” in Proceedings of the 22nd International Society for Music Information Re- trieval Conference, ISMIR, 2021, pp. 88–96

    2021

  43. [51]

    MelodyT5: A unified score-to-score transformer for symbolic mu- sic processing,

    S. Wu, Y . Wang, X. Li, F. Yu, and M. Sun, “MelodyT5: A unified score-to-score transformer for symbolic mu- sic processing,” in Proceedings of the 25th Interna- tional Society for Music Information Retrieval Confer- ence, ISMIR, 2024, pp. 642–650

    2024

  44. [52]

    Compound word transformer: Learning to compose full-song mu- sic over dynamic directed hypergraphs,

    W. Hsiao, J. Liu, Y . Yeh, and Y . Yang, “Compound word transformer: Learning to compose full-song mu- sic over dynamic directed hypergraphs,” in Thirty- Fifth AAAI Conference on Artificial Intelligence, AAAI, 2021, pp. 178–186

    2021

  45. [53]

    Attention is all you need,

    A. Vaswani et al., “Attention is all you need,” in Ad- vances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems, 2017, pp. 5998–6008

    2017

  46. [54]

    Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory ef- ficient, and long context finetuning and inference,

    B. Warner et al. , “Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory ef- ficient, and long context finetuning and inference,” CoRR, vol. abs/2412.13663, 2024

    2024 arXiv

  47. [55]

    FlashAttention: fast and memory-efficient exact at- tention with IO-awareness,

    T. Dao, D. Y . Fu, S. Ermon, A. Rudra, and C. Ré, “FlashAttention: fast and memory-efficient exact at- tention with IO-awareness,” in Advances in Neural Information Processing Systems 35: Annual Confer- ence on Neural Information Processing Systems 2022, NeurIPS, 2022

    2022

  48. [56]

    Roformer: Enhanced transformer with ro- tary position embedding,

    J. Su, M. H. M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: Enhanced transformer with ro- tary position embedding,” Neurocomputing, vol. 568, p. 127063, 2024

    2024

  49. [57]

    Representation deficiency in masked language modeling,

    Y . Meng et al., “Representation deficiency in masked language modeling,” inThe Twelfth International Con- ference on Learning Representations, ICLR, 2024

    2024

  50. [58]

    Perfor- mance midi-to-score conversion by neural beat track- ing,

    L. Liu, Q. Kong, V . Morfi, and E. Benetos, “Perfor- mance midi-to-score conversion by neural beat track- ing,” in Proceedings of the 23rd International Society for Music Information Retrieval Conference, ISMIR , 2022, pp. 395–402

    2022

  51. [59]

    POP909: A pop-song dataset for mu- sic arrangement generation,

    Z. Wang et al., “POP909: A pop-song dataset for mu- sic arrangement generation,” inProceedings of the 21th International Society for Music Information Retrieval Conference, ISMIR 2020, 2020, pp. 38–45

    2020

  52. [60]

    A cor- pus describing orchestral texture in first movements of classical and early-romantic symphonies,

    D. Le, M. Giraud, F. Levé, and F. Maccarini, “A cor- pus describing orchestral texture in first movements of classical and early-romantic symphonies,” in DLfM ’22: 9th International Conference on Digital Libraries for Musicology, 2022, pp. 27–35

    2022

  53. [61]

    Orchestral texture classification with convolution,

    Y .-H. Chu and L. Su, “Orchestral texture classification with convolution,” in Extended Abstracts for the Late- Breaking Demo Session of the 24th International So- ciety for Music Information Retrieval Conference (IS- MIR), 2023

    2023

  54. [62]

    Position- ing left-hand movement in violin performance: A sys- tem and user study of fingering pattern generation,

    Y .-H. Jen, T.-P. Chen, S.-W. Sun, and L. Su, “Position- ing left-hand movement in violin performance: A sys- tem and user study of fingering pattern generation,” in Proceedings of the 26th International Conference on Intelligent User Interfaces, 2021, p. 208–212

    2021

  55. [63]

    Enhancing violin fin- gering generation through audio-symbolic fusion,

    W. Lin, Y . F. Wang, and L. Su, “Enhancing violin fin- gering generation through audio-symbolic fusion,” in IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP, 2024, pp. 811–815

    2024

  56. [64]

    Stable and low- precision training for large-scale vision-language mod- els,

    M. Wortsman, T. Dettmers, L. Zettlemoyer, A. Mor- cos, A. Farhadi, and L. Schmidt, “Stable and low- precision training for large-scale vision-language mod- els,” in Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems...

    2023

  57. [65]

    Learning-based methods for comparing se- quences, with applications to audio-to-midi alignment and matching,

    C. Raffel, “Learning-based methods for comparing se- quences, with applications to audio-to-midi alignment and matching,” Ph.D. dissertation, Columbia Univer- sity, USA, 2016

    2016

  58. [66]

    Multi-mmlg: a novel framework of ex- tracting multiple main melodies from MIDI files,

    J. Zhao, D. Taniar, K. Adhinugraha, V . M. Baskaran, and K. Wong, “Multi-mmlg: a novel framework of ex- tracting multiple main melodies from MIDI files,”Neu- ral Comput. Appl., vol. 35, no. 30, pp. 22 687–22 704, 2023

    2023

  59. [67]

    The Llama 3 herd of models,

    A. Dubey et al., “The Llama 3 herd of models,”CoRR, vol. abs/2407.21783, 2024

    2024 arXiv

  60. [68]

    Deep learning scaling is predictable, empirically,

    J. Hestness et al., “Deep learning scaling is predictable, empirically,”CoRR, vol. abs/1712.00409, 2017

    2017 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.