Pith. sign in

REVIEW 2 major objections 4 minor 42 references

This paper argues that symbolic music generation should be reframed as explicit editing of a draft, and that the reason edit-based methods never transferred from text to music is the encoding: one note scattered across interleaved event tok

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 01:55 UTC pith:5D2O6ZLK

load-bearing objection Worth engaging: a genuinely new framing with solid internal results, but the synthetic-data circularity means the headline advantage over AR/diffusion is not yet proven for real edits. the 2 major comments →

arxiv 2607.11124 v2 pith:5D2O6ZLK submitted 2026-07-13 cs.SD cs.AI

BeatEdit: Symbolic Music Generation as Explicit Editing

Haoyu Gu , Lekai Qian , Haowu Zhou , Qi Liu , Shuai Wang This is my paper
classification cs.SD cs.AI
keywords symbolic music generationedit-based generationmusic representationbeat-grid encodingsequence taggingiterative refinementtag-then-fillnon-autoregressive generation
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.

BeatEdit tries to establish that symbolic music generation is better understood as editing a draft than as generating a score from scratch. The paper traces the absence of edit-based music methods to the representation: in conventional event encodings a single note is splintered across several interleaved tokens, so edits cascade and nothing aligns. It adopts the BEAT encoding, which anchors each beat to a fixed grid slot and compresses each pitch's within-beat rhythm into a single pattern token, giving each note a self-contained unit that source and target can align beat-for-beat. On this base it builds three mechanisms for increasing edit density—per-token tagging, iterative delete-insert-fill, and tag-then-fill—sharing one pre-trained bidirectional encoder, and reports they nearly double the best generative baseline in exact-match correction while running in tens of milliseconds. The paper also claims encoding choice is a major design lever, with relative pitch offsets helping single-pass tagging but hurting iterative methods.

Core claim

The paper's central claim is that edit-based generation works for symbolic music once the representation satisfies three structural requirements: each note is an atomic token group, source and target align position-by-position, and a local edit stays local. Conventional encodings fail these requirements; the beat-grid-anchored BEAT encoding satisfies them by packing each beat into fixed token spans and each pitch's rhythm into one base-3 pattern token. Given such an encoding, three mechanisms on a shared masked-language-model encoder—sequence tagging for sparse corrections, iterative deletion-insertion-prediction for accompaniment rewriting, and tag-then-fill for segment completion—localize

What carries the argument

The load-bearing object is the BEAT encoding, a beat-grid anchored tokenization. A three-state piano roll is cut into beat columns; for each active pitch, its within-beat rhythm is compressed by base-3 into one pattern token (81 possible values), and pitches are laid out in sorted order as position tokens, with beats interleaved across voices. This gives every note a self-contained token pair (or one bundled token), makes the n-th beat of source and target occupy the same span regardless of edits, and confines changes to a beat, satisfying the paper's R1-R3 requirements. The residual dependency—relative pitch offsets that cascade when one pitch changes—is handled by a music-specific SHIFT op

Load-bearing premise

The load-bearing premise is that the synthetic corruptions used to create draft-final pairs—pitch shifts, rhythm changes, note deletions and insertions at fixed probabilities—faithfully stand in for the way people actually revise music; if real editing involves structural changes outside those perturbation types, the reported advantage of BeatEdit may not transfer.

What would settle it

Collect a corpus of genuine human revisions (an earlier draft and the finalized score of the same piece), run BeatEdit on those pairs, and check whether exact-match recovery and human ratings still beat autoregressive and diffusion baselines. A simpler probe: take a clean score, apply edits not representable as per-beat perturbations—moving a phrase by a half-bar, changing time signature, merging voices, or deleting an entire section—and see whether the editing mechanisms recover the target; the current alignment is beat-count preserving and would be expected to fail on such inputs by construc

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

If this is right

  • If the paper is right, symbolic music models can correct, rearrange, and fill without regenerating: the editor localizes positions to change and applies typed operations, preserving everything else verbatim.
  • The encoding—not just the model—determines whether editing works; a fixed beat grid and atomic per-note tokens become design requirements, and relative pitch offsets should be avoided in multi-pass editing pipelines.
  • The three mechanisms span an edit-density axis: sparse typo-style fixes, dense accompaniment rewrites, and segment completion share one backbone, and the axis has a principled end—when no draft survives, editing reverts to generation and precision collapses.
  • Single-pass editing is not just convenient: roughly 1-3 forward passes versus thousands of autoregressive decoding steps makes editing two orders of magnitude faster, a gap the paper argues is architectural rather than implementation detail.
  • Pre-training a bidirectional encoder on music is the largest single driver; removing it drops correction accuracy by 72%, and an encoder-decoder variant collapses, implying shared attention between edited and context tokens is essential.

Where Pith is reading between the lines

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

  • Beyond the paper, the synthetic-corruption protocol means the reported doubling is a measurement of recovery of the corrupted target under the same perturbation distribution. A natural next test is a real paired corpus of human revisions, which the paper itself calls for, to see if the advantage survives structural edits like re-barring, tempo changes, or arrangement-level rewrites.
  • Beyond the paper, the edit-density axis suggests a continuum rather than a dichotomy: an edit-based model with an empty draft reduces to generation, so the framework could subsume 'generate from scratch' as the dense extreme, and the finding that distributional quality collapses exactly when the last source token disappears gives a precise place to switch modes.
  • Beyond the paper, the SHIFT adaptation for relative offsets is a general lesson for edit-based modeling of any sequence with translational-invariance encodings, such as code diffs or time-series; the paper's encoding-method interaction warns that a representation that is good for prediction can be bad for editing.
  • Beyond the paper, the POP909 comparison hints at a hybrid: editing methods produce more precise content while diffusion produces more natural onset timing; combining an editor that fixes exact content with a diffusion stage that re-times it could beat either alone.

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

2 major / 4 minor

Summary. The paper introduces BeatEdit, a framework that reformulates symbolic music generation as explicit editing of a draft, using a beat-grid-anchored encoding (Beat) that satisfies three structural requirements for edit-based methods (R1–R3). Three mechanisms are proposed along an edit-density axis: SeqTag for sparse correction, IterEdit for iterative accompaniment editing, and TagFill for segment completion, all sharing a pre-trained BERT-style encoder. The paper reports that edit-based methods outperform autoregressive, diffusion, and inpainting baselines on synthetic correction/editing/completion tasks, with a 2×2 encoding-factor analysis showing that encoding choice strongly interacts with method choice. Extensive experiments include 192K piano pieces, bootstrap CIs, an external POP909 comparison, MOS studies, multi-track transfer, and latency/FLOP analyses.

Significance. If the empirical claims hold, the paper makes a useful conceptual contribution by importing explicit edit operations from NLP into symbolic music generation and by identifying representation-level requirements for editing. The R1–R3 formalization is clean, and the cross-encoding factorial is a valuable design analysis. The paper is also unusually thorough on evaluation infrastructure: held-out splits, bootstrap CIs, ablations, MOS, and an external comparison with a released baseline. The main risk is external validity: the tasks are defined by a synthetic corruption protocol, and the central claim of superiority over general-purpose generative baselines may be specific to that protocol.

major comments (2)
  1. [§4.1, Appendix G; Table 4] The principal empirical claim rests on a single synthetic protocol: edit pairs are generated by corrupting clean targets with the perturbation types and probabilities of Appendix G, and all main-task metrics are computed on the same protocol. Because BeatEdit's modules are trained to invert exactly this corruption process, the near-doubling over generative baselines in Table 4 may reflect specialization to the training perturbation distribution rather than a general editing capability. The §4.7 boundary experiment varies only the amount of remaining context, not the type of edit; real edits such as bar insertion/deletion, reharmonization, or tempo/key changes violate the beat-count-preserving assumptions and are untested. Since the abstract and conclusion state an unqualified advantage 'across all three tasks,' this is a load-bearing gap. I would ask for either (i) an evaluation on natur
  2. [§4.3–§4.4; Tables 4 and 17] The headline comparisons in Table 4 select the best encoding scheme per method–task cell, but Table 17 shows the choice is consequential and method-dependent: e.g., IterEdit correction is 0.719 on Scheme A but 0.070 on B; TagFill editing is 0.350 on D but 0.089 on B; SeqTag correction peaks on B (0.726) but does not use the same scheme for the other methods. The significant encoding×method interactions (§M.2) mean the reported 'single encoding and pre-trained backbone' framework does not yield the headline numbers with one fixed encoding; a practitioner must know the task-specific optimal scheme. To make the central 'higher precision across all three tasks' claim actionable, I recommend reporting a fixed-encoding comparison (e.g., the same scheme for all methods, or a principled scheme-selection rule) in addition to the best-per-cell numbers.
minor comments (4)
  1. [Abstract; Appendix O] The abstract claims 'single-pass inference completing in under 100 ms,' but IterEdit's editing latency is 655 ms and completion is 82 ms (Table 27). Please qualify the claim as applying to the single-pass mechanisms (SeqTag, TagFill) or to the typical case, not to all BeatEdit variants.
  2. [§4.2 vs. Table 4] §4.2 states No-Edit baselines have note_f1 ≈ 0.881–0.884, while Table 4 reports No-Edit note_f1 = 0.644 for correction on perturbed beats only. Since §4.1 says all metrics except FMD are computed on perturbed beats, the two numbers appear inconsistent. Please clarify which beat subset is used in each table.
  3. [Table 3; Appendix J] The 'recovery' metric in Table 3 is described only in Appendix J. Please define it in the main text (even parenthetically) so the cross-encoding comparison is self-contained.
  4. [§2.3, Table 2] The table's 'Avg. len' column reports sequence length, but the text later reports token lengths in the range 1,200–2,500 with different values. Please reconcile the numbers or clarify the exact setting (e.g., dual-voice vs. single-track, max tokens).

Circularity Check

0 steps flagged

No significant circularity: central claims rest on held-out synthetic evaluation plus external POP909 baselines; the self-defined corruption protocol is a limitation, not a reduction.

full rationale

The paper's derivation chain is not circular. The three editing mechanisms are defined directly on the Beat token structure (Section 3), and their training targets are generated by synthetic perturbation of clean targets (Section 4.1, Appendix G). The headline comparisons (Table 4) are measured on held-out test pieces under the same perturbation protocol, but the test set is disjoint from training and the models must generalize to unseen content; no fitted parameter is renamed as a prediction. The encoding's structural advantages (R1-R3) are tested empirically against four alternative encodings under the same SeqTag architecture (Section 4.2), so the choice of Beat is not justified solely by self-citation. The external POP909 comparison against Polyffusion (Table 6) provides independent benchmark support. The main honest limitation, stated in Section 4.1 and the Conclusion, is that synthetic corruptions may not capture real-world editing diversity; this affects external validity and generalizability, but it is not a circularity because the evaluation does not reduce to the training input by construction. The self-citation to the authors' BEAT encoding [28] is load-bearing but is supported by in-paper cross-encoding experiments, making it independent evidence rather than a circular appeal.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on a self-defined synthetic task and on the BEAT representation from the authors' prior work. The main free parameters are perturbation probabilities, difficulty mixture, and gating thresholds; no new physical or representational entities are postulated.

free parameters (5)
  • Synthetic perturbation probabilities = pitch shift 0.10, rhythm 0.05, deletion 0.03, insertion 0.02, no change 0.80
    Chosen ad hoc in Appendix G; defines the task distribution on which all main results are measured.
  • Difficulty level mixture = L1:L2:L3:L4 = 30:30:25:15
    Used in training and evaluation (§4.1, D.2); directly shapes the aggregate scores in Table 4.
  • SeqTag gating hyperparameters = KEEP weight 0.15, lambda_detect 0.5, error threshold theta 0.5, KEEP bias beta 0.3
    Tuned hyperparameters controlling the conservative bias; not derived from first principles.
  • SHIFT range = ±1..±5 semitones
    Introduced in §3.2 to handle relative-encoding cascades; chosen by hand.
  • Time resolution tau = 4 steps per beat
    Fixed quantization resolution for all experiments (§2.2); hand-chosen, affects token vocabulary and edit granularity.
axioms (4)
  • domain assumption BEAT encoding satisfies R1–R3 for editing
    Paper relies on the structural properties of the BEAT representation from [28] (same authors); Table 1 is an informal comparison, not a proof. Section 2.2.
  • domain assumption Synthetic corruptions approximate real music edits
    Section 4.1 and the conclusion explicitly acknowledge the field lacks paired draft-to-final corpora; all main-task results are on this synthetic protocol.
  • domain assumption Exact-match recovery on perturbed beats is a valid measure of editing quality
    Section 4.1 defines metrics; the paper notes it is blind to plausible alternatives, and adds distribution metrics only in the POP909 comparison.
  • domain assumption BERT MLM pre-training on 192K piano pieces transfers to edit tasks
    Used as the shared backbone; ablation shows it is critical (§4.7, L.2), but no theory is given for why it transfers so strongly.

pith-pipeline@v1.3.0-alltime-deepseek · 26571 in / 11408 out tokens · 124634 ms · 2026-08-03T01:55:02.293087+00:00 · methodology

0 comments
read the original abstract

Music creation is fundamentally a process of revision. Yet symbolic music generation remains dominated by paradigms that produce complete sequences from scratch, with limited support for selective modification. Edit-based methods have proven effective for text transformation tasks, but remain largely unexplored for symbolic music. We trace this absence to the representational level: conventional event-based music encodings lack the structural properties required by explicit music editing. In contrast, the BEAT encoding, a beat-grid-anchored representation originally designed for autoregressive generation, possesses structural properties amenable to editing. We propose BeatEdit, the first framework for symbolic music generation based on explicit edit operations, recasting generation as producing new content by editing a draft rather than synthesizing from scratch. BeatEdit comprises three complementary mechanisms along an axis of increasing edit density: per-token sequence tagging for error correction, iterative refinement for accompaniment editing, and tag-then-fill for segment completion. All these mechanisms share a single encoding and pre-trained backbone, achieving higher precision and perceptual quality than autoregressive and diffusion methods across all three tasks, while remaining efficient, with single-pass inference completing in under 100 ms. Cross-encoding evaluation further reveals that encoding design substantially influences editing effectiveness, with notable encoding-method interaction effects. Code is available at https://github.com/Haoyu-Gu/BeatEdit-code

Figures

Figures reproduced from arXiv: 2607.11124 by Haowu Zhou, Haoyu Gu, Lekai Qian, Qi Liu, Shuai Wang.

Figure 1
Figure 1. Figure 1: Encoding pipeline. Each pitch becomes a pattern token [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: IterEdit pipeline. Stage 1: BERT MLM pre-training. Stage 2: three heads trained jointly with the encoder. Stage 3: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: IterEdit pipeline. Stage 1: BERT MLM pre-training. Stage 2: three heads trained jointly with the encoder. Stage 3: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SeqTag correction Edit F1 by error type across en [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: SeqTag correction Edit F1 by error type across en [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SeqTag correction Edit F1 by sequence length across [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: SeqTag correction Edit F1 by sequence length across [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

Review history (2 revisions) →

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 · 6 linked inside Pith

  1. [1]

    Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

    Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. 2021. Structured Denoising Diffusion Models in Discrete State-Spaces. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 34

    2021

  2. [2]

    Abhijeet Awasthi, Sunita Sarawagi, Rasna Goyal, Sabyasachi Ghosh, and Vihari Piratla. 2019. Parallel Iterative Edit Models for Local Sequence Transduction. In BeatEdit: Symbolic Music Generation as Explicit Editing Conference’17, July 2017, Washington, DC, USA Proceedings of the 2019 Conference on Empirical Methods in Natural Language Pro- cessing and the...

    2019

  3. [3]

    Chin-Jui Chang, Chun-Yi Lee, and Yi-Hsuan Yang. 2021. Variable-Length Mu- sic Score Infilling via XLNet and Musically Specialized Positional Encoding. InProceedings of the 22nd International Society for Music Information Retrieval Conference (ISMIR). 97–104

    2021

  4. [4]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). Association for Comput...

    2019

  5. [5]

    Elizabeth Dinella, Hanjun Dai, Ziyang Li, Mayur Naik, Le Song, and Ke Wang

  6. [6]

    Hao-Wen Dong, Ke Chen, Shlomo Dubnov, Julian McAuley, and Taylor Berg- Kirkpatrick. 2023. Multitrack Music Transformer. InICASSP 2023 – IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 1–5

    2023

  7. [7]

    Nathan Fradet, Jean-Pierre Briot, Fabien Chhel, Amal El Fallah Seghrouchni, and Nicolas Gutowski. 2023. MidiTok: A Python Package for MIDI File Tokenization. arXiv preprint arXiv:2310.17202(2023)

    Pith/arXiv arXiv 2023

  8. [8]

    Nathan Fradet, Nicolas Gutowski, Fabien Chhel, and Jean-Pierre Briot. 2023. Im- pact of Time and Note Duration Tokenizations on Deep Learning Symbolic Music Modeling. InProceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR). 89–97

    2023

  9. [9]

    Jiatao Gu, Changhan Wang, and Jake Zhao. 2019. Levenshtein Transformer. In Advances in Neural Information Processing Systems, Vol. 32

    2019

  10. [10]

    Rui Guo, Ivor Simpson, Chris Kiefer, Thor Magnusson, and Dorien Herremans

  11. [11]

    Gaëtan Hadjeres and Léopold Crestel. 2021. The Piano Inpainting Application. arXiv preprint arXiv:2107.05944(2021)

    Pith/arXiv arXiv 2021

  12. [12]

    Wen-Yi Hsiao, Jen-Yu Liu, Yin-Cheng Yeh, and Yi-Hsuan Yang. 2021. Compound Word Transformer: Learning to Compose Full-Song Music over Dynamic Directed Hypergraphs. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 178–186

    2021

  13. [13]

    Dai, Matthew D

    Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne, Noam Shazeer, Andrew M. Dai, Matthew D. Hoffman, Monica Din- culescu, and Douglas Eck. 2019. Music Transformer: Generating Music with Long-Term Structure. InInternational Conference on Learning Representations (ICLR). arXiv:1809.04281

    Pith/arXiv arXiv 2019

  14. [14]

    Yujia Huang, Adishree Ghatare, Yuanzhe Liu, Ziniu Hu, Qinsheng Zhang, Chan- dramouli Shama Sastry, Siddharth Gururani, Sageev Oore, and Yisong Yue. 2024. Symbolic Music Generation with Non-Differentiable Rule Guided Diffusion. In Proceedings of the 41st International Conference on Machine Learning (ICML) (Pro- ceedings of Machine Learning Research, Vol. 2...

    2024

  15. [15]

    Yu-Siang Huang and Yi-Hsuan Yang. 2020. Pop Music Transformer: Beat-based Modeling and Generation of Expressive Pop Piano Compositions. InProceedings of the 28th ACM International Conference on Multimedia (MM ’20). ACM, 1180–1188. doi:10.1145/3394171.3413671

    arXiv 2020

  16. [16]

    Wengong Jin, Kevin Yang, Regina Barzilay, and Tommi Jaakkola. 2019. Learning Multimodal Graph-to-Graph Translation for Molecular Optimization. InInterna- tional Conference on Learning Representations (ICLR)

    2019

  17. [17]

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Fo- cal Loss for Dense Object Detection. InIEEE International Conference on Computer Vision (ICCV). 2980–2988

    2017

  18. [18]

    Peiling Lu, Xin Xu, Chenfei Kang, Botao Yu, Chengyi Xing, Xu Tan, and Jiang Bian. 2023. MuseCoco: Generating Symbolic Music from Text.arXiv preprint arXiv:2306.00110(2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Ang Lv, Xu Tan, Peiling Lu, Wei Ye, Shikun Zhang, Jiang Bian, and Rui Yan. 2023. GETMusic: Generating Any Music Tracks with a Unified Representation and Diffusion Framework.arXiv preprint arXiv:2305.10841(2023)

    Pith/arXiv arXiv 2023

  20. [20]

    Jonathan Mallinson, Aliaksei Severyn, Eric Malmi, and Guillermo Garrido. 2020. FELIX: Flexible Text Editing Through Tagging and Insertion. InFindings of the Association for Computational Linguistics: EMNLP 2020. Association for Computa- tional Linguistics, 1244–1255

    2020

  21. [21]

    Eric Malmi, Sebastian Krause, Sascha Rothe, Daniil Mirylenka, and Aliaksei Severyn. 2019. Encode, Tag, Realize: High-Precision Text Editing. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP- IJCNLP). Association for Computational Lingui...

    2019

  22. [22]

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. 2022. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. InInternational Conference on Learning Representations (ICLR)

    2022

  23. [23]

    Lejun Min, Junyan Jiang, Gus Xia, and Jingwei Zhao. 2023. Polyffusion: A Diffusion Model for Polyphonic Score Generation with Internal and External Controls. InProceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR). 231–238

    2023

  24. [24]

    Kostiantyn Omelianchuk, Vitaliy Atrasevych, Artem Chernodub, and Oleksandr Skurzhanskyi. 2020. GECToR – Grammatical Error Correction: Tag, Not Rewrite. InProceedings of the Fifteenth Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2020). Association for Computational Linguistics, 163–170

    2020

  25. [25]

    Sageev Oore, Ian Simon, Sander Dieleman, Douglas Eck, and Karen Simonyan

  26. [26]

    Ashis Pati, Alexander Lerch, and Gaëtan Hadjeres. 2019. Learning to Traverse Latent Spaces for Musical Score Inpainting. InProceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR). 343–351

    2019

  27. [27]

    Christine Payne. 2019. MuseNet. OpenAI Blog. https://openai.com/index/ musenet/

    2019

  28. [28]

    doi:10.1007/s00521-018-3758-9

    This Time with Feeling: Learning Expressive Musical Performance.Neural Computing and Applications32 (2020), 955–967. doi:10.1007/s00521-018-3758-9

    doi:10.1007/s00521-018-3758-9 2020

  29. [29]

    2016.Learning-Based Methods for Comparing Sequences, with Appli- cations to Audio-to-MIDI Alignment and Matching

    Colin Raffel. 2016.Learning-Based Methods for Comparing Sequences, with Appli- cations to Audio-to-MIDI Alignment and Matching. Ph. D. Dissertation. Columbia University

    2016

  30. [30]

    Vipul Raheja, Dhruv Kumar, Ryan Koo, and Dongyeop Kang. 2023. CoEdIT: Text Editing by Task-Specific Instruction Tuning. InFindings of the Associa- tion for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics, 5274–5291

    2023

  31. [31]

    Lekai Qian, Haoyu Gu, Jingwei Zhao, and Ziyu Wang. 2026. BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps. InInternational Conference on Machine Learning (ICML)

    2026

  32. [32]

    John Thickstun, David Hall, Chris Donahue, and Percy Liang. 2024. Anticipa- tory Music Transformer. InInternational Conference on Learning Representations (ICLR)

    2024

  33. [33]

    Dimitri von Rütte, Luca Biggio, Yannic Kilcher, and Thomas Hofmann. 2023. FIGARO: Controllable Music Generation using Learned and Expert Features. In International Conference on Learning Representations (ICLR)

    2023

  34. [34]

    Jan Retkowski, Jakub Stępniak, and Mateusz Modrzejewski. 2024. Fréchet Music Distance: A Metric for Generative Symbolic Music Evaluation. arXiv:2412.07948 [cs.SD] https://arxiv.org/abs/2412.07948

    Pith/arXiv arXiv 2024

  35. [35]

    Ziyu Wang, Lejun Min, and Gus Xia. 2024. Whole-Song Hierarchical Generation of Symbolic Music Using Cascaded Diffusion Models. InProceedings of the Twelfth International Conference on Learning Representations (ICLR)

    2024

  36. [36]

    Li-Chia Yang and Alexander Lerch. 2020. On the Evaluation of Generative Models in Music.Neural Computing and Applications32 (2020), 4773–4784

    2020

  37. [37]

    Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, and Sheng Zhao. 2023. AUDIT: Audio Editing by Following Instructions with La- tent Diffusion Models. InAdvances in Neural Information Processing Systems 36 (NeurIPS)

    2023

  38. [38]

    Huan Zhang, Emmanouil Karystinaios, Simon Dixon, Gerhard Widmer, and Carlos Eduardo Cancino-Chacón. 2023. Symbolic Music Representations for Clas- sification Tasks: A Systematic Evaluation. InProceedings of the 24th International Society for Music Information Retrieval Conference (ISMIR). 848–858

    2023

  39. [39]

    full” variants use the multi-level perturbation strategy (Section 3.3 of the main paper); “vanilla

    Yixiao Zhang, Yukara Ikemiya, Gus Xia, Naoki Murata, Marco A. Martínez- Ramírez, Wei-Hsiang Liao, Yuki Mitsufuji, and Simon Dixon. 2024. MusicMagus: Zero-Shot Text-to-Music Editing via Diffusion Models. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence (IJCAI). 7805– 7813. Conference’17, July 2017, Washington, DC,...

    2024

  40. [40]

    Mingliang Zeng, Xu Tan, Rui Wang, Zeqian Ju, Tao Qin, and Tie-Yan Liu. 2021. MusicBERT: Symbolic Music Understanding with Large-Scale Pre-Training. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021. Asso- ciation for Computational Linguistics, 791–800

    2021

  41. [2020]

    InInternational Conference on Learning Representations (ICLR)

    Hoppity: Learning Graph Transformations to Detect and Fix Bugs in Programs. InInternational Conference on Learning Representations (ICLR)

  42. [2022]

    InArtificial Intelligence in Music, Sound, Art and Design (EvoMUSART 2022) (Lecture Notes in Computer Science)

    MusIAC: An Extensible Generative Framework for Music Infilling Applica- tions with Multi-level Control. InArtificial Intelligence in Music, Sound, Art and Design (EvoMUSART 2022) (Lecture Notes in Computer Science). Springer, 341–356

    2022