Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

ImprovNet -- Generating Controllable Musical Improvisations with Iterative Corruption Refinement

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A single transformer turns classical pieces into jazz, harmonizes melodies, and infills gaps.

desk verdict Solid system paper with a human-verified genre-conversion result, but the full-piece structural-coherence claim lacks support and needs referee pressure. read the letter →

arxiv 2502.04522 v4 pith:FCMB3YLO submitted 2025-02-06 cs.SD cs.AIeess.AS

Keshav Bhandari , Sungkyun Chang , Tongyu Lu , Fareza R. Enus , Louis B. Bradshaw , Dorien Herremans , Simon Colton This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords musicgenerationstyletransferimprovisationcorruptionrefinementtransformersymbolicharmonizationinfilling
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

This paper claims that a single transformer, trained by corrupting short segments of piano pieces and teaching the model to repair them, can act as a unified engine for expressive music generation: it can re-improvise a complete classical or jazz piece in another genre, rewrite it within its own genre, harmonize a given melody, and continue or infill short musical prompts. The reason this matters is that these tasks are normally handled by separate specialised models, and symbolic music style transfer for complete works has been hindered by scarce jazz data and rigid, quantized representations. ImprovNet addresses both by using weak genre labels and a minimally quantized tokenizer, and by making the corruption functions themselves the user's control knobs: choosing corruption type, corruption rate, and number of passes governs how much of the original structure survives. If the claim holds, a single model offers user-tunable style transfer for full compositions rather than fragments.

What carries the argument

The engine is the corruption-refinement loop over 5-second segments. A 5-second chunk is corrupted by one of nine functions (pitch/velocity masking, onset/duration masking, whole masking, pitch permutation, fragmentation, incorrect transposition, note modification, skyline melody extraction), and the transformer encoder-decoder must reconstruct the original segment from its left and right context, conditioned on genre and corruption-type tokens. Repeated passes over the whole piece, with a corruption rate and a preservation ratio that keeps novel segments intact, gradually reshape the music toward the target genre. This same machinery serves all four tasks: full-piece style transfer, harmonization (skyline corruption constrains the first chord of each segment), and short continuation/infilling (whole-mask corruption with trimmed context).

What would settle it

Take a corpus of full-length classical pieces, run classical-to-jazz conversion with the recommended settings, and after each pass compute the chroma-based SSM correlation and a long-range theme-repetition metric between the original and the generated piece; if the SSM correlation decays to chance level after a few passes while local 5-second segment statistics remain plausible, the claim that complete-piece structure is preserved is falsified.

Watch

Extended reading notes

Core claim

The central claim is that iterative corruption refinement unifies multiple music-generation capabilities in one architecture. The model is trained to reconstruct a corrupted 5-second segment from its surrounding context, conditioned on a genre token and a corruption-type token; at generation time, the user repeatedly corrupts and refines every segment of a complete piece, with a target genre token, for several passes. Each pass pushes the piece gradually toward the target genre, and by choosing which corruption functions to apply, at what rate, and how many segments to preserve, the user controls the trade-off between stylistic transformation and structural similarity to the original. The paper reports that this procedure yields recognizable classical-to-jazz conversion (79% of listeners identified the jazz condition), favourable objective and subjective results against the Anticipatory Music Transformer on short continuation and infilling, and genre-conditioned harmonization of monophonic melodies via a skyline corruption plus logit constraints.

Load-bearing premise

The load-bearing premise is that repeatedly corrupting and refining 5-second segments, with at most 55 seconds of surrounding context, preserves the global structure of a complete multi-minute piece rather than drifting over many passes; the paper's own note that coherence deteriorates beyond 20 seconds for continuation shows where this assumption is most vulnerable.

Editorial extensions

If this is right

  • A single ImprovNet model can perform cross-genre and intra-genre improvisation, harmonization, short continuation, and short infilling without task-specific fine-tuning architectures.
  • Users can control the intensity of style transfer and structural similarity by adjusting corruption functions, corruption rates, number of passes, and the preservation ratio, rather than by retraining the model.
  • ImprovNet outperforms the Anticipatory Music Transformer baseline on 10-second continuation and infilling measured by pitch-class KL divergence, pitch-class transition cosine similarity, note density, and inter-onset interval.
  • Because corruption refinement acts as data augmentation, the model can be pre-trained on large classical corpora and fine-tuned on a smaller jazz corpus, easing the data bottleneck for jazz style transfer.
  • Harmonization with logit constraints produces dense, genre-specific chords for monophonic melodies, achieving 76% correct identification of jazz harmonizations.

Reading between the lines

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

  • An implicit corollary: because the generation procedure is compositional over segments and corruption functions, the framework could be ported to other autoregressive sequence domains (text, code, other symbolic music formats) by defining domain-appropriate corruptions and genre/attribute tokens, without changing the training objective.
  • The paper's own admission that coherence deteriorates beyond 20 seconds in continuation suggests a natural stress test: measuring structural similarity after many passes on pieces longer than the 55-second training context would reveal whether the iterative loop converges to a stylistically consistent whole or drifts into local incoherence.
  • A testable extension would be ablating the preservation ratio systematically: the current experiments fix it at 0.05, so varying it while holding corruption functions and passes fixed would map the user-control surface more completely.
  • Because the skyline corruption already extracts a monophonic melody, the harmonization task could be inverted into a melody-extraction evaluation: feed the model a polyphonic piece, extract the skyline, harmonize it, and check how close the reharmonization comes to the original harmony.
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 / 6 minor

Summary. The paper proposes ImprovNet, an encoder-decoder transformer trained with a self-supervised corruption-refinement strategy on expressive solo-piano MIDI. A complete piece is divided into 5-second segments; during training a random segment is corrupted and the model refines it back given surrounding context and a genre token; during generation, segments are iteratively corrupted and refined over multiple passes with user-specified corruption functions, corruption rates, and a preservation ratio. The authors claim this unifies cross-genre and intra-genre improvisation, harmonization, short continuation, and infilling in a single model, with user control over the degree of style transfer and structural similarity. Evaluations include a self-trained genre classifier, SSM-based structural similarity, objective metrics against Anticipatory Music Transformer, and listening tests with 28 participants. Code and a demo page are provided.

Significance. If the central claims are supported, ImprovNet would be a useful contribution to expressive symbolic music generation: it addresses a real gap by unifying multiple tasks and offers a plausible mechanism for controllable style conversion. The paper has concrete strengths: it ships code and a demo page, the corruption-refinement formulation is clearly described, and the subjective evaluation provides independent evidence for genre conversion (79% identification, one-sample binomial test p = 0.0037) and a listener preference signal over the AMT baseline. However, the headline claim about complete multi-minute works with preserved global structure rests on an untested assumption about iterative local refinement, and the principal objective genre metric is partly self-referential. These issues are load-bearing for the central contribution and require additional evidence or a narrowed claim.

major comments (3)
  1. [Section III-F1 and Eq. (6)] The central claim of style-aware improvisations for complete musical works assumes that iteratively refining local 5-second segments, each conditioned on at most 11 segments (55 seconds) of context, preserves global structure over a multi-minute piece. This assumption is not established. Training (Fig. 2) corrupts one random segment at a time, and no training objective or analysis addresses convergence, error accumulation, or long-range thematic and sectional coherence over Q passes. The paper itself notes in Section III-F2 that coherency deteriorates beyond 20 seconds for continuation, which is exactly the regime needed for complete-piece CGI/IGI. The supporting evidence (SSM correlation curves in Figs. 4b and 5, and subjective structure ratings in Tables IV-V) is aggregated and lacks per-piece distributions, error bars, and any direct test of section order or repetition. I ask the authors to provide a systematic analysis of structural drift over passes - for example, section-level segmentation precision/recall, repetition-order metrics, or per-piece SSM distributions on the test set - or to narrow the scope of the complete-works claim accordingly.
  2. [Section IV-B2 and Section V-A1] The objective evidence for cross-genre conversion relies on a genre classifier trained on the same fine-tuning datasets (Maestro, PiJAMA, and Doug McKenzie) and the same Aria tokenization as the generator. Because the classifier has effectively been exposed to the same genre definitions and data as the model, the 'cross-genre probability' increases in Figs. 4-5 may partly reflect the classifier's familiarity with training-data characteristics rather than an independent measurement of genre transfer. This is not fatal because the listening test provides independent support, but the metric should be validated by reporting classifier accuracy on held-out external data, by using a pre-trained external genre classifier, or by explicitly acknowledging and quantifying the circularity.
  3. [Tables I and II] Tables I and II report single point estimates for each metric on the short continuation and infilling tasks, with no error bars, confidence intervals, or significance tests, despite the text claiming 'superior performance' over AMT. If these numbers are averages over multiple generated examples, per-example variance and paired tests (e.g., Wilcoxon signed-rank) should be reported; otherwise differences such as PCTM 0.347 versus 0.307 in Table I may be within noise. The human preference result in Fig. 6 is helpful, but the objective claim needs proper statistical reporting to be convincing.
minor comments (6)
  1. [Section VI-B] In the paragraph after Fig. 6, 'Finally, table 6 based on Section 4' should refer to Fig. 6, not Table 6.
  2. [Section III-C, Eq. (6)] Equation (6) as written is ambiguous: r_theta is defined in Eq. (5) as acting on the corrupted input sequence Sc,g with context, but the recurrence in Eq. (6) shows r_theta applied directly to the previous segment sr(q-1) without explicitly describing the corruption step and context construction. Please rewrite the recurrence to make the corrupted input explicit.
  3. [Section III-C] The condition 'Pq < alpha' overloads P as both the pass index and a random variable; please clarify that a random draw is compared with the corruption rate alpha at each segment and pass.
  4. [Section IV-B2] For the genre classifier, please report its own held-out accuracy or F1 score and specify the input representation and decision threshold; this is needed to interpret the classifier-based objective results.
  5. [Section V-A1] The SSM correlation is computed from 'the audio of the original and generated compositions', but the generation pipeline is symbolic; please specify the synthesis/rendering software and settings used to convert MIDI to audio, since this affects chroma-based structural similarity.
  6. [Tables IV and V] The 'Human-like score is assumed as 5 for the original' convention should be described as an anchoring ceiling, not an experimental measurement, and should not be directly compared with the model scores without a caveat.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ImprovNet's generation and evaluation chain is self-contained; auxiliary genre classifier is self-referential but not a fitted-input prediction.

full rationale

ImprovNet's central generation pipeline is self-contained: the corruption-refinement training objective (Eq. 5) and iterative generation procedure (Eq. 6) are fully specified in this paper, and no parameter of the generator is fitted to the evaluation targets. The genre classifier used in Section VI-A is trained on the same classical and jazz data and shares the encoder architecture, which makes that objective metric partly self-referential, but it is an evaluation instrument rather than a fitted input to the generator, and the paper's subjective listening test (79% genre identification, p = 0.0037) provides independent support for the cross-genre claim. The self-citations to prior work by the authors ([10], [32], [44]) are not load-bearing: the corruption-refinement strategy is described in full in this manuscript, and no uniqueness theorem or external mathematical result is imported to force the model choice. The acknowledged 20-second coherence limit in Section III-F2 is a limitation on the full-piece structural claim, but it is a correctness risk rather than a circular reduction. No equation reduces to its own input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central claim rests on hand-chosen corruption parameters (alpha, context lengths, pass count, chord window), on the reliability of automatically transcribed jazz data and genre labels, and most importantly on the unvalidated assumption that local iterative refinement preserves global musical structure. No new physical entities are introduced; the genre token and corruption functions are procedural, not ontological additions.

free parameters (8)
  • corruption_rate alpha = 0.25-1.0 in experiments; user-set
    Controls the fraction of segments refined per pass; Fig. 5 shows genre probability and structural similarity depend on it.
  • preservation ratio = 0.05 in objective experiments
    Fraction of original segments kept intact; hand-set value affects structural similarity and user control.
  • number of refinement passes Q = up to 10 in Fig. 4
    User-set; more passes increase genre probability and decrease structural similarity in the reported curves.
  • left and right context lengths L,R = 1-5 during training; not fixed at generation
    Controls the local context window; authors note CGI works better with small right context, but no exact generation setting is reported.
  • temperature tau = not reported
    Sampling temperature in Equation 8; no value given, yet it affects the stochasticity of every generated output.
  • harmonization chord onset window = 50 ms
    Equation 9 constrains chord onsets within 50 ms of the first melody note; hand-chosen value affects chord alignment.
  • corruption function parameters = e.g., fragmentation 20-50%, transposition +/-5 semitones, note modification 10-40%
    Hand-chosen ranges for corruption functions defined in Section III-D; these determine the training signal and the type of stylistic change.
  • segment chunk length = 5000 ms
    Aria tokenizer chunk length; controls vocabulary size and the unit of corruption and refinement.
assumptions (7)
  • standard math Transformer encoder-decoder attention captures musical dependencies from tokenized segments.
    The model's capacity to reconstruct corrupted segments relies on this unproved background assumption about attention and the transformer inductive bias.
  • domain assumption The Aria tokenizer's chunked absolute-onset encoding preserves enough expressive performance information for genre transfer and harmonization.
    Section III-A; the approach encodes timing, dynamics, and pedal, but it also discards bars, key, and time signatures, and quantization may lose information.
  • domain assumption Automatic retranscription of PiJAMA audio yields jazz MIDI accurate enough for fine-tuning.
    Section IV-A; retranscription via an external AMT model can introduce errors that affect what the model learns about jazz style.
  • domain assumption The genre labels on the datasets are reliable and each piece can be assigned to a single target genre.
    Training and evaluation condition on a binary classical or jazz token; mixed-genre or mislabeled pieces would weaken genre conditioning.
  • domain assumption The nine corruption functions create a learnable mapping from corrupted fragments back to originals that supports stylistic modification.
    Section III-D; if the corruption functions do not align with how jazz and classical differ, the model learns repair rather than style transfer.
  • ad hoc to paper Iterative refinement of 5-second segments with limited context preserves global structure over complete pieces.
    This is the load-bearing design assumption for whole-piece CGI and IGI; Section III-F2 notes coherency deteriorates beyond 20 seconds for continuation, and no direct analysis of long-range structural drift is provided.
  • ad hoc to paper The genre classifier trained on the same data provides a valid objective measure of genre conversion.
    The classifier is trained on the same classical and jazz datasets as the generator, so its probability partially reflects statistical familiarity rather than independent genre perception.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ImprovNet -- Generating Controllable Musical Improvisations with Iterative Corruption Refinement." pith.science (2026). https://pith.science/paper/FCMB3YLO

@misc{pith2026250204522,
  author       = {Pith},
  title        = {Pith review of: ImprovNet -- Generating Controllable Musical Improvisations with Iterative Corruption Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCMB3YLO}},
  note         = {Machine review of arXiv:2502.04522}
}
read the original abstract

Despite deep learning's remarkable advances in style transfer across various domains, generating controllable performance-level musical style transfer for complete symbolically represented musical works remains a challenging area of research. Much of this is owed to limited datasets, especially for genres such as jazz, and the lack of unified models that can handle multiple music generation tasks. This paper presents ImprovNet, a transformer-based architecture that generates expressive and controllable musical improvisations through a self-supervised corruption-refinement training strategy. The improvisational style transfer is aimed at making meaningful modifications to one or more musical elements - melody, harmony or rhythm of the original composition with respect to the target genre. ImprovNet unifies multiple capabilities within a single model: it can perform cross-genre and intra-genre improvisations, harmonize melodies with genre-specific styles, and execute short prompt continuation and infilling tasks. The model's iterative generation framework allows users to control the degree of style transfer and structural similarity to the original composition. Objective and subjective evaluations demonstrate ImprovNet's effectiveness in generating musically coherent improvisations while maintaining structural relationships with the original pieces. The model outperforms Anticipatory Music Transformer in short continuation and infilling tasks and successfully achieves recognizable genre conversion, with 79\% of participants correctly identifying jazz-style improvisations of classical pieces. Our code and demo page can be found at https://github.com/keshavbhandari/improvnet.

Figures

Figures reproduced from arXiv: 2502.04522 by the authors.

Figure 1
Figure 1. A 10 second piano roll depicting the Aria tokenizer that uses absolute onsets relative to the start of the segment. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. A single pass iterates over all the segments in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. A) A random segment from S is selected with context segments around it. B) This segment is corrupted with a corruption function. C) ImprovNet refines it back. B. ImprovNet Overview ImprovNet relies on a corruption-refinement training strat￾egy similar to [44], to generate music in an improvisational style of the original composition. The entire piece is segmented into 5-second segments as seen in [PITH_FULL_IMAGE:f… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Cross-genre improvisations generated with all corruption functions using a corruption rate of 1.0. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Effect of various corruption rates on cross-genre [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Musicality preference comparison among AMT, Im [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RUMAA: Repeat-Aware Unified Music Audio Analysis for Score-Performance Alignment, Transcription, and Mistake Detection

    cs.SD 2025-07 conditional novelty 6.0 of 10

    A single transformer model aligns scores to performances, transcribes piano audio, and detects mistakes, including faithful handling of repeat sections without pre-unfolded scores.

Reference graph

Works this paper leans on

52 extracted references · 38 canonical work pages · cited by 1 Pith paper

  1. [1]

    Image style transfer using convolutional neural networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Image style transfer using convolutional neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 2414–2423

    2016

  2. [2]

    Unsupervised melody style conversion,

    E. Nakamura, K. Shibata, R. Nishikimi, and K. Yoshii, “Unsupervised melody style conversion,” in ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019

    work page 2019

  3. [3]

    Supervised Symbolic Music Style Translation Using Synthetic Data

    O. C ´ıfka, U. S ¸ims ¸ekli, and G. Richard, “Supervised symbolic music style translation using synthetic data,” arXiv preprint:1907.02265, 2019

    work page Pith review arXiv 1907

  4. [4]

    Many-to- many symbolic multi-track music genre transfer,

    M. Pezzat, H. Perez-Meana, T. Nakashika, and M. Nakano, “Many-to- many symbolic multi-track music genre transfer,” in Knowledge Innova- tion Through Intelligent Software Methodologies, Tools and Techniques. IOS Press, 2020, pp. 272–281

    work page 2020

  5. [5]

    Unpaired image- to-image translation using cycle generative adversarial networks [j],

    A. Dwarkani, M. Jain, J. Thakkar, and K. Kottursamy, “Unpaired image- to-image translation using cycle generative adversarial networks [j],” International Journal of Engineering and Advanced Technology (IJEAT), vol. 9, no. 6, 2020. 4https://stable-diffusion-art.com/inpainting basics/

    work page 2020

  6. [6]

    Composeinstyle: Music composition with and without style transfer,

    S. Mukherjee and M. Mulimani, “Composeinstyle: Music composition with and without style transfer,” Expert Systems with Applications , vol. 191, p. 116195, 2022

    work page 2022

  7. [7]

    Groove2groove: One-shot music style transfer with supervision from synthetic data,

    O. C ´ıfka, U. S ¸ims ¸ekli, and G. Richard, “Groove2groove: One-shot music style transfer with supervision from synthetic data,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 28, pp. 2638–2650, 2020

    work page 2020

  8. [8]

    Music style transfer: A position paper,

    S. Dai, Z. Zhang, and G. G. Xia, “Music style transfer: A position paper,” arXiv preprint arXiv:1803.06841 , 2018

    arXiv 2018

Show all 52 references
  1. [9]

    Pijama: Piano jazz with automatic midi annotations,

    D. Edwards, S. Dixon, and E. Benetos, “Pijama: Piano jazz with automatic midi annotations,” Transactions of the International Society for Music Information Retrieval , 2023

    2023

  2. [10]

    Motifs, phrases, and beyond: The modelling of structure in symbolic music generation,

    K. Bhandari and S. Colton, “Motifs, phrases, and beyond: The modelling of structure in symbolic music generation,” in International Conference on Computational Intelligence in Music, Sound, Art and Design (Part of EvoStar). Springer, 2024, pp. 33–51

    2024

  3. [11]

    R. L. Crocker, A history of musical style . Courier Corporation, 1986

    1986

  4. [12]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  5. [13]

    Musemorphose: Full-song and fine-grained piano music style transfer with one transformer vae,

    S.-L. Wu and Y .-H. Yang, “Musemorphose: Full-song and fine-grained piano music style transfer with one transformer vae,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 31, pp. 1953– 1967, 2023

    1953

  6. [14]

    Deepbach: a steerable model for bach chorales generation,

    G. Hadjeres, F. Pachet, and F. Nielsen, “Deepbach: a steerable model for bach chorales generation,” in International conference on machine learning. PMLR, 2017, pp. 1362–1371

    2017

  7. [15]

    Counterpoint by convolution,

    C.-Z. A. Huang, T. Cooijmans, A. Roberts, A. Courville, and D. Eck, “Counterpoint by convolution,” arXiv preprint arXiv:1903.07227, 2019

    1903 arXiv

  8. [16]

    When counterpoint meets chinese folk melodies,

    N. Jiang, S. Jin, Z. Duan, and C. Zhang, “When counterpoint meets chinese folk melodies,” Advances in neural information processing systems, vol. 33, pp. 16 258–16 270, 2020

    2020

  9. [17]

    Music fadernets: Controllable music generation based on high-level features via low-level feature modelling,

    H. H. Tan and D. Herremans, “Music fadernets: Controllable music generation based on high-level features via low-level feature modelling,” in ISMIR, 2020

    2020

  10. [18]

    Accomontage: Accompaniment arrangement via phrase selection and style transfer,

    J. Zhao and G. Xia, “Accomontage: Accompaniment arrangement via phrase selection and style transfer,” arXiv preprint:2108.11213, 2021

    2021 arXiv

  11. [19]

    Generating chord progression from melody with flexible harmonic rhythm and controllable harmonic density,

    S. Wu, Y . Yang, Z. Wang, X. Li, and M. Sun, “Generating chord progression from melody with flexible harmonic rhythm and controllable harmonic density,” EURASIP Journal on Audio, Speech, and Music Processing, vol. 2024, no. 1, p. 4, 2024

    2024

  12. [20]

    Getmusic: Generating any music tracks with a unified representation and diffusion framework,

    A. Lv, X. Tan, P. Lu, W. Ye, S. Zhang, J. Bian, and R. Yan, “Getmusic: Generating any music tracks with a unified representation and diffusion framework,” arXiv preprint arXiv:2305.10841 , 2023

    2023 arXiv

  13. [21]

    Popmag: Pop music accompaniment generation,

    Y . Ren, J. He, X. Tan, T. Qin, Z. Zhao, and T.-Y . Liu, “Popmag: Pop music accompaniment generation,” in Proceedings of the 28th ACM international conference on multimedia , 2020, pp. 1198–1206

    2020

  14. [22]

    Structured multi-track accompaniment arrangement via style prior modelling,

    J. Zhao, G. Xia, Z. Wang, and Y . Wang, “Structured multi-track accompaniment arrangement via style prior modelling,” in The 38th Annual Conference on Neural Information Processing Systems , 2024

    2024

  15. [23]

    C2-magic: Chord-controllable multi-track ac- companiment generation with interpretability and creativity,

    J. Wu, Z. Ji, and P. Li, “C2-magic: Chord-controllable multi-track ac- companiment generation with interpretability and creativity,” in Summit on Music Intelligence . Springer, 2023, pp. 108–121

    2023

  16. [24]

    Polyffusion: A diffusion model for polyphonic score generation with internal and external controls,

    L. Min, J. Jiang, G. Xia, and J. Zhao, “Polyffusion: A diffusion model for polyphonic score generation with internal and external controls,” arXiv preprint arXiv:2307.10304 , 2023

    2023 arXiv

  17. [25]

    Performance rnn: Generating music with expressive timing and dynamics,

    I. Simon and S. Oore, “Performance rnn: Generating music with expressive timing and dynamics,” https://magenta.tensorflow.org/ performance-rnn, 2017

    2017

  18. [26]

    Graph neural network for music score data and modeling expressive piano performance,

    D. Jeong, T. Kwon, Y . Kim, and J. Nam, “Graph neural network for music score data and modeling expressive piano performance,” in ICML. PMLR, 2019, pp. 3060–3070

    2019

  19. [27]

    Virtuosonet: A hier- archical rnn-based system for modeling expressive piano performance

    D. Jeong, T. Kwon, Y . Kim, K. Lee, and J. Nam, “Virtuosonet: A hier- archical rnn-based system for modeling expressive piano performance.” in ISMIR, 2019, pp. 908–915

    2019

  20. [28]

    Music performance style transfer for learning expressive musical performance,

    Z. Xiao, X. Chen, and L. Zhou, “Music performance style transfer for learning expressive musical performance,” Signal, Image and Video Processing, vol. 18, no. 1, pp. 889–898, 2024

    2024

  21. [29]

    Scoreperformer: Expressive piano performance rendering with fine-grained control

    I. Borovik and V . Viro, “Scoreperformer: Expressive piano performance rendering with fine-grained control.” in ISMIR, 2023, pp. 588–596

    2023

  22. [30]

    Rendering music perfor- mance with interpretation variations using conditional variational rnn

    A. Maezawa, K. Yamamoto, and T. Fujishima, “Rendering music perfor- mance with interpretation variations using conditional variational rnn.” in ISMIR, 2019, pp. 855–861

    2019

  23. [31]

    Pertok: Expressive encoding and mod- eling of symbolic musical ideas and variations,

    J. Lenz and A. Mani, “Pertok: Expressive encoding and mod- eling of symbolic musical ideas and variations,” arXiv preprint arXiv:2410.02060, 2024

    2024 arXiv

  24. [32]

    Automatic generation of expressive piano miniatures

    S. Colton, L. Bradshaw, B. Banar, K. Bhandari et al. , “Automatic generation of expressive piano miniatures.” International Conference on Computational Creativity (ICCC), 2024

    2024

  25. [33]

    Symbolic music genre transfer with cyclegan,

    G. Brunner, Y . Wang, R. Wattenhofer, and S. Zhao, “Symbolic music genre transfer with cyclegan,” in 2018 ieee 30th international conference on tools with artificial intelligence (ictai) . IEEE, 2018, pp. 786–793

    2018

  26. [34]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2223–2232

    2017

  27. [35]

    Transfer symbolic music style from latent representation,

    Y . Fu, Y . Tanimura, and H. Nakada, “Transfer symbolic music style from latent representation,” in Proceedings of the 34th Annual Conference of the Japanese Society for Artificial Intelligence (2020) . The Japanese Society for Artificial Intelligence, 2020, pp. 2K4ES201–2K4ES201

    2020

  28. [36]

    beta-vae: Learning basic vi- sual concepts with a constrained variational framework

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner, “beta-vae: Learning basic vi- sual concepts with a constrained variational framework.” ICLR (Poster), vol. 3, 2017

    2017

  29. [37]

    Genre style transfer : Symbolic genre style transfer utilising gan with additional genre-enforcing discriminators,

    L. Sulaiman and S. Larsson, “Genre style transfer : Symbolic genre style transfer utilising gan with additional genre-enforcing discriminators,” Master’s thesis, Halmstad University, School of Information Technology, 2022

    2022

  30. [38]

    Steelygan: semantic un- supervised symbolic music genre transfer,

    Z. Ding, X. Liu, G. Zhong, and D. Wang, “Steelygan: semantic un- supervised symbolic music genre transfer,” in Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 2022, pp. 305–317

    2022

  31. [39]

    Anticipatory music transformer,

    J. Thickstun, D. Hall, C. Donahue, and P. Liang, “Anticipatory music transformer,” arXiv preprint arXiv:2306.08620 , 2023

    2023 arXiv

  32. [40]

    Variable-length music score infilling via xlnet and musically specialized positional encoding,

    C.-J. Chang, C.-Y . Lee, and Y .-H. Yang, “Variable-length music score infilling via xlnet and musically specialized positional encoding,” arXiv preprint arXiv:2108.05064, 2021

    2021 arXiv

  33. [41]

    The piano inpainting application,

    G. Hadjeres and L. Crestel, “The piano inpainting application,” arXiv preprint arXiv:2107.05944, 2021

    2021 arXiv

  34. [42]

    What algorithms can transformers learn? a study in length generalization,

    H. Zhou, A. Bradley, E. Littwin, N. Razin, O. Saremi, J. Susskind, S. Bengio, and P. Nakkiran, “What algorithms can transformers learn? a study in length generalization,” arXiv preprint arXiv:2310.16028, 2023

    2023 arXiv

  35. [43]

    Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions,

    Y .-S. Huang and Y .-H. Yang, “Pop music transformer: Beat-based modeling and generation of expressive pop piano compositions,” in Proceedings of the 28th ACM international conference on multimedia , 2020, pp. 1180–1188

    2020

  36. [44]

    Yin-yang: Developing motifs with long-term structure and controllability,

    K. Bhandari, G. A. Wiggins, and S. Colton, “Yin-yang: Developing motifs with long-term structure and controllability,” arXiv preprint arXiv:2501.17759, 2025

    2025 arXiv

  37. [45]

    Self-similarity-based and novelty-based loss for music structure analysis,

    G. Peeters, “Self-similarity-based and novelty-based loss for music structure analysis,” in Proceedings of the 24th International Society for Music Information Retrieval Conference, ISMIR 2023 , 2023

    2023

  38. [46]

    Melodic matching techniques for large music databases,

    A. Uitdenbogerd and J. Zobel, “Melodic matching techniques for large music databases,” in Proceedings of the seventh ACM international conference on Multimedia (Part 1) , 1999, pp. 57–66

    1999

  39. [47]

    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 ISMIR, 2022, pp. 757–763

    2022

  40. [48]

    Atepp: A dataset of automatically transcribed expressive piano performance,

    H. Zhang, J. Tang, S. R. M. Rafee, and S. D. G. Fazekas, “Atepp: A dataset of automatically transcribed expressive piano performance,” in ISMIR 2022 Hybrid Conference , 2022

    2022

  41. [49]

    Enabling factorized piano music modeling and generation with the maestro dataset,

    C. Hawthorne, A. Stasyuk, A. Roberts, I. Simon, C.-Z. A. Huang, S. Dieleman, E. Elsen, J. Engel, and D. Eck, “Enabling factorized piano music modeling and generation with the maestro dataset,” arXiv preprint arXiv:1810.12247, 2018

    arXiv

  42. [50]

    Symbolic music similarity through a graph-based representation,

    F. Simonetta, F. Carnovalini, N. Orio, and A. Rod `a, “Symbolic music similarity through a graph-based representation,” in Proceedings of the Audio Mostly 2018 on Sound in Immersion and Emotion , 2018, pp. 1–7

    2018

  43. [51]

    On the evaluation of generative models in music,

    L.-C. Yang and A. Lerch, “On the evaluation of generative models in music,” Neural Comput. Appl. , vol. 32, no. 9, p. 4773–4784, May

  44. [2020]

    Available: https://doi.org/10.1007/s00521-018-3849-7

    [Online]. Available: https://doi.org/10.1007/s00521-018-3849-7

    doi

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.