Pith. sign in

REVIEW 5 major objections 6 minor 40 references

RenderBox: Expressive Performance Rendering with Text Control

T0 review · 5 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 RenderBox is a diffusion-transformer system that renders a symbolic MIDI score into audio performances whose tempo, expressiveness, mistakes, and style are directed by natural-language prompts, trained with a curriculum that builds from…

desk verdict RenderBox is a credible first step in text-and-score expressive audio rendering, but the 'unified framework' claim outruns the evidence because no single checkpoint is shown to retain all controls. read the letter →

arxiv 2502.07711 v1 pith:T7GNN4HB submitted 2025-02-11 eess.AS cs.MM

Huan Zhang , Akira Maezawa , Simon Dixon This is my paper · ORCID
classification eess.AScs.MM
keywords expressiveperformancerenderingtext-to-audiodiffusiontransformerMIDIconditioningcurriculumlearningcontrolmulti-instrumentmusicgenerationmistakesimulation
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 introduces RenderBox, a system that turns a symbolic MIDI score plus a short text description into an audio performance. The authors claim it is the first model to support text-based control over expressive performance attributes such as tempo, playing style, mistakes, and emotional nuance, and to do so across piano, guitar, saxophone, violin, and orchestral instruments. The proposed framework applies coarse controls through natural language and fine-grained controls through the score, using a diffusion transformer with cross-attention joint conditioning and a curriculum that progresses from plain synthesis to style-directed performance. If the claim holds, the same score can be rendered as a neutral playback, a speed-varied performance, a mistake-prone student take, or a performance in a named pianist's style, all from one model.

What carries the argument

The central mechanism is a joint conditioning scheme in which MIDI event tokens, T5 text embeddings, and a duration embedding are concatenated along the sequence dimension and fed to the cross-attention layers of a diffusion transformer (DiT) operating on a 21.5 Hz latent audio representation. The MIDI tokens use the instrument, note, on/off, time, and end-of-tie vocabulary from MT3, which lets the model read a 10-second polyphonic score window. The second load-bearing component is a curriculum learning scheduler that orders training data from strict synthesis to expressive and style-directed performance, with speed augmentation and a mistake-simulation algorithm generating the intermediate stages. This staged schedule is what the paper credits for allowing the model to acquire fine-grained controls while retaining the score-following ability.

What would settle it

Render the same MIDI score with a single model trained through all five stages and ask a listener (or a pitch-detection system) whether the output still matches the score's pitches and tempo when prompted with 'synthesis'; if the final model cannot reproduce stage-0 accuracy, the unified-framework claim fails.

Watch

Extended reading notes

Core claim

RenderBox is presented as the first unified framework that generates expressive musical audio conditioned jointly by a MIDI score and a natural-language prompt. The model is built on the Stable Audio Open diffusion-transformer backbone, with the MIDI score tokenized using the MT3 vocabulary and injected together with text and duration embeddings through cross-attention. The authors propose a five-stage curriculum that starts with pure MIDI-to-audio synthesis, adds tempo control, then expressive performances, then mistake-included performances, and finally style-directed performances labeled by performer or expression. Objective metrics (FAD, CLAP, pitch and tempo accuracy) and a subjective listening test are reported to show that RenderBox outperforms audio-generation baselines and matches symbolic performance-rendering systems while adding text-based flexibility. The authors also report an interpretable embedding space in which generations cluster by performer and composer.

Load-bearing premise

The framework's claim of being a unified controller for all five task types assumes that a single set of weights retains its stage-0 synthesis ability after training on later, more variable performance stages; the paper's own section 5.3 reports that stage-2 and stage-4 checkpoints lose synthesis performance, so the claim rests on retention that is only partially demonstrated.

Editorial extensions

If this is right

  • The same score can be rendered as neutral synthesis, a speed-changed performance, an expressive interpretation, a student-like take with mistakes, or a performance in a specified pianist's style, all by one model with text prompts.
  • Text prompts such as 'a bit slower, expressive performance, Bach, Piano' directly control tempo tier, performance type, composer, and instrumentation, with common terminology mapped to quantitative tempo ranges.
  • The curriculum schedule (synthesis to speed to expression to mistakes to style) provides a general recipe for incrementally teaching a generative model new control dimensions.
  • Compared with symbolic performance-rendering systems, RenderBox outputs audio directly rather than MIDI, and it is the only tested model capable of text-specified speed control.

Reading between the lines

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

  • The authors evaluate different checkpoints for different stages rather than a single model run through the full curriculum; if a final model is expected to handle all five task types at once, the reported forgetting of stage-0 synthesis would need to be addressed, for instance by rehearsal or regularization.
  • The speed and mistake controls could be extended to other instruments not in the training set, such as brass or organ, provided aligned score-performance data exist; the architecture itself is instrument-agnostic in its MIDI conditioning.
  • The style embedding space could allow zero-shot style interpolation between performers if the learned factors are truly linear, a test the paper does not run.
  • The mistake-corruption algorithm is synthetic; a natural next step would be to train on real human mistakes and evaluate whether listeners perceive the degradation as authentic rather than mechanical.
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

5 major / 6 minor

Summary. The paper introduces RenderBox, a text-and-score conditioned audio performance rendering system built on a Stable Audio Open diffusion transformer with MIDI-token conditioning, and proposes a five-stage curriculum ranging from plain synthesis through speed-controlled synthesis, expressive performance, mistake-corrupted performance, and style-directed performance. The authors evaluate the system against MusicGen-Melody, Coco-Mulla, MIDI-DDSP, and symbolic performance-rendering baselines using FAD, CLAP, pitch accuracy, tempo deviation, and a 23-participant subjective study, and they report a t-SNE analysis of performer and composer embedding structure. The manuscript claims that a single RenderBox framework achieves flexible text-based control for speed, mistakes, and style diversity across multiple instruments.

Significance. If the framework performs as claimed, it would be a meaningful step toward controllable expressive audio rendering: one model that maps a symbolic score and a natural-language prompt to synthesized, speed-varied, expressive, mistake-including, or style-directed audio. The paper's strengths include the broad aggregation of publicly available aligned score-performance datasets, the explicit curriculum design with staged control factors, and the combination of objective audio metrics with a human listening study that reports statistically significant advantages over several baselines. The t-SNE analysis also provides an interesting view of learned performer and composer structure. However, the central claim of a single unified model is not established by the reported evidence: Table 3 evaluates separate checkpoints for different stages, §5.3 explicitly documents forgetting across stages, and the mistake-conditioned stage is not objectively evaluated. The contribution is potentially significant, but the manuscript currently overstates the demonstrated capabilities.

major comments (5)
  1. [§5.3, Table 3, §6] The conclusion states that RenderBox 'achieves flexible control for speed, mistakes, and style diversity with multiple instruments,' but Table 3 evaluates only stage-0, stage-2, and stage-4 checkpoints (RenderBox-Stage0, RenderBox-Stage2, RenderBox-Stage4), and §5.3 explicitly says that models trained up to stages 2 and 4 lose stage-0 synthesis performance. No experiment demonstrates that one final checkpoint retains all skills simultaneously. Please either provide a single-checkpoint evaluation across all five curriculum stages, or qualify the central claim to describe a curriculum-trained family of models rather than one unified model.
  2. [§5.1, Table 3, §4.1] Mistake control is one of the three headline controls listed in the abstract and conclusion, yet Table 3 contains no stage-3 row and the objective metrics (FAD, CLAP, pitch, tempo) are not reported for mistake-corrupted outputs. The evidence for mistake control is limited to a spectrogram example in Figure 4 and a pooled subjective score in Figure 3. Without a per-stage objective evaluation or a significance-tested subjective comparison specifically targeting the mistake condition, the claim of flexible mistake control is unsupported. Please add a stage-3 evaluation or temper the mistake-control claim.
  3. [§4, §5.4, §6] The style-direction stage (stage 4) is trained exclusively on piano datasets (ATEPP, Con Espressione, Vienna 4x22, as listed in Table 1), and the t-SNE analysis in §5.4 uses piano outputs only. Consequently, the statement that RenderBox achieves 'style diversity with multiple instruments' is not demonstrated for guitar, saxophone, violin, or orchestral instruments. Please either evaluate style transfer for non-piano instruments or restrict the claim to piano-style control.
  4. [§5.1, Table 3] The objective results are reported as point estimates without confidence intervals, error bars, or significance tests, and it is unclear whether the evaluations were run over multiple seeds or independent generation sets. Since the paper uses these numbers to claim that RenderBox 'largely outperforms' baselines, the absence of statistical characterization is load-bearing. Please report confidence intervals or repeated-run variability, and state whether the observed differences are significant.
  5. [§5.2, Figure 3] The subjective evaluation is described with very limited protocol detail: 23 participants, nine questions, six instruments, and pooled MOS dimensions, with a single p-value threshold mentioned. The report does not state how many stimuli were rated per condition, how the stimuli were selected, whether listening was controlled for audio quality or presentation order, or whether multiple-comparison corrections were applied. Given that the 'natural and musically engaging' claim rests substantially on this study, please provide a fuller protocol description and per-condition results.
minor comments (6)
  1. [Abstract] The phrase 'capture musical can emotional intent' appears to contain a typo; it should likely read 'capture musical and emotional intent.'
  2. [§5.3] The section contains the placeholder text '[to add]' and the following analysis sentence is incomplete; this section must be finished before publication.
  3. [§4, Table 1] The in-house saxophone dataset is described as copyrighted and is not released, which limits reproducibility; please state its availability status or describe a publicly available alternative that could be used for reproduction.
  4. [Table 3] The table omits the stage-3 condition entirely; if stage-3 results are added, the table should include a corresponding row to make the stage-wise evaluation complete.
  5. [Table 2] The speed prompt list includes paraphrases such as 'Twice as slow as' and 'About half the speed of' that denote the same tier; consider stating explicitly whether these are treated as synonymous during training.
  6. [Figure 4] The caption says the visualizations are 20-second windows, while the model input window is described as 10 seconds in §3.1; please clarify how these windows were constructed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RenderBox's claims rest on directly supervised conditional mappings evaluated with external audio metrics, not on definitional reductions.

full rationale

RenderBox is an empirical systems paper: its central claims concern training a conditional diffusion model, not deriving a result from first principles. The speed-control evaluation reuses the training label vocabulary ('a bit faster') as test prompts, but that is the task definition, and the tempo check is computed with an external estimator (madmom) against the score tempo rather than by reading back the training label. Style control is evaluated with CLAP text-alignment and with t-SNE visualizations on a test subset; performer-ID and expression labels are conditioning inputs, not the output of the evaluation. Cited self-works (ATEPP, the mistake taxonomy, DExter) supply datasets, augmentation recipes, and baselines; none is invoked as a uniqueness theorem or as a premise that already contains the conclusion. The forgetting result in Section 5.3 and the piano-only stage-4 data are real limitations of the evidence relative to the Section 6 claim (one checkpoint retaining all controls; style diversity across instruments), but they are coverage and overclaim issues, not circular reductions. No equation or metric in the paper is identical by construction to its input, and no fitted parameter is renamed as a prediction. The appropriate finding is therefore no significant circularity.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on transfer from the pretrained base model, the tokenization choice, the curriculum schedule, and the validity of the evaluation metrics. No new physical or conceptual entities are introduced. The paper does not release code or weights, and the forgetting result means the 'unified' final model is not demonstrated to retain all stage capabilities.

free parameters (5)
  • Curriculum stage step counts = 20k / 10k / 15k / 4k / 10k steps
    Hand-selected in Section 3.2; no ablation of the schedule except the no-CL mixed-data condition. The central controllability result depends on this schedule.
  • Speed tier boundaries and keyword map = Duration ratios 0.4-0.6, 0.6-0.8, 0.8-1.2, 1.2-1.5, 1.5-1.8, 1.8-2.2 with fixed keywords
    Defined in Table 2; these ranges define the speed-control task and the tempo-deviation evaluation, but no sensitivity analysis is provided.
  • Mistake augmentation probabilities = 5% mistouch, 20% asynchrony, 5% substitution, 5% ghost note; block removal every 5s
    Given in Algorithm 1; the mistake-control capability is trained and evaluated using these ad hoc probabilities without ablations.
  • Classifier-free guidance scale = 7.0 with DPM-Solver++ for 100 steps
    Inference setting in Section 3.1; affects prompt adherence and audio quality, but is not ablated.
  • DTW chroma penalty weight lambda = 1e-3
    Introduced in Section 5.1 as a weighting factor for the DTW alignment cost in pitch accuracy; no sensitivity analysis is reported.
assumptions (6)
  • domain assumption The pre-trained Stable Audio Open autoencoder, T5 text embedder, and DiT provide a viable base for score-conditioned music generation after fine-tuning.
    Invoked in Section 3.1; if the pretrained latent space does not support long, polyphonic score-faithful rendering, the whole curriculum result collapses.
  • domain assumption MT3 MIDI-like tokenization of the input score preserves all information needed to render the correct notes and instrumentation.
    Used in Section 3.1; tokenization may lose velocity curves, pedal, and other expressive details that could limit faithful rendering.
  • domain assumption The curriculum order (synthesis, speed, expressive, mistakes, style) is what produces the reported gains, and the final stage model retains earlier capabilities.
    Section 3.2 and the ablation in Section 5.3; however, the paper reports forgetting of stage 0 in later checkpoints, so the single-model retention part is questionable.
  • domain assumption Objective audio metrics (FAD on OpenL3, CLAP text-output distance, DTW chroma similarity, madmom tempo deviation) are valid proxies for score fidelity, prompt alignment, and tempo control.
    Used in Section 5.1; no calibration or error bars are provided, so the scores may not fully reflect perceived quality.
  • domain assumption The public datasets and their alignments ((n)ASAP, MusicNet, GAPS, ATEPP, Con Espressione, Vienna 4x22) are accurate enough for supervised score-to-performance training.
    Listed in Table 1; alignment errors would bias both training targets and evaluation ground truths.
  • domain assumption Performer IDs and expression labels from ATEPP and Con Espressione can serve as style descriptions in natural-language prompts.
    Stage 4 uses performer names and expression labels as text conditioning; this assumes these labels capture separable style information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RenderBox: Expressive Performance Rendering with Text Control." pith.science (2026). https://pith.science/paper/T7GNN4HB

@misc{pith2026250207711,
  author       = {Pith},
  title        = {Pith review of: RenderBox: Expressive Performance Rendering with Text Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7GNN4HB}},
  note         = {Machine review of arXiv:2502.07711}
}
read the original abstract

Expressive music performance rendering involves interpreting symbolic scores with variations in timing, dynamics, articulation, and instrument-specific techniques, resulting in performances that capture musical can emotional intent. We introduce RenderBox, a unified framework for text-and-score controlled audio performance generation across multiple instruments, applying coarse-level controls through natural language descriptions and granular-level controls using music scores. Based on a diffusion transformer architecture and cross-attention joint conditioning, we propose a curriculum-based paradigm that trains from plain synthesis to expressive performance, gradually incorporating controllable factors such as speed, mistakes, and style diversity. RenderBox achieves high performance compared to baseline models across key metrics such as FAD and CLAP, and also tempo and pitch accuracy under different prompting tasks. Subjective evaluation further demonstrates that RenderBox is able to generate controllable expressive performances that sound natural and musically engaging, aligning well with prompts and intent.

Figures

Figures reproduced from arXiv: 2502.07711 by the authors.

Figure 1
Figure 1. An overview of the performance space proposed by our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. ControlNet conditioning (left) and concatenative cross [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. MOS score of the subjective evaluation on the four dimen [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Input MIDI piano rolls and output spectrograms with respect to different text prompting. All visualizations are 20-second windows. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of generation with testing data subset, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Cail- lon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, Matt Sharifi, Neil Zeghidour, and Christian Frank

    [Agostinelli et al., 2023] Andrea Agostinelli, Timo I. Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Cail- lon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, Matt Sharifi, Neil Zeghidour, and Christian Frank. MusicLM: Generating Music From Text. arXiv preprint arXiv:2301.11325,

    arXiv 2023

  2. [4]

    Director Musices : The KTH Performance Rules System

    [Bresin et al., 2002] Roberto Bresin, Anders Friberg, and Jo- han Sundberg. Director Musices : The KTH Performance Rules System. Special Interest Group on Music and Com- puter(SIGMUS) - 46 Kyoto, pages 43–48,

    work page 2002

  3. [9]

    Look, Listen, and Learn More: Design Choices for Deep Audio Embeddings,

    [Cramer et al., 2019] Jason Cramer, Ho Hsiang Wu, Justin Salamon, and Juan Pablo Bello. Look, Listen, and Learn More: Design Choices for Deep Audio Embeddings,. In In Proceedings of the ICASSP 2019 - 2019 IEEE In- ternational Conference on Acoustics, Speech and Signal (ICASSP),

    work page 2019

  4. [12]

    From MNIST to ImageNet and back: benchmarking continual curriculum learning

    [Faber et al., 2024] Kamil Faber, Dominik Zurek, Marcin Pietron, Nathalie Japkowicz, Antonio Vergari, and Roberto Corizzo. From MNIST to ImageNet and back: benchmarking continual curriculum learning. Machine Learning, 113(10), mar

    work page 2024

  5. [13]

    Ren- con: Performance Rendering Contest for Automated Mu- sic Systems

    [Hashida et al., 2008] M Hashida, M Nakra, H Katayose, T Murao, K Hirata, K Suzuki, and T Kitahara. Ren- con: Performance Rendering Contest for Automated Mu- sic Systems. In Proceedings of the 10th International Conference on Music Perception and Cognition (ICMPC)., Sapporo,

    work page 2008

  6. [15]

    VirtuosoNet: A Hierar- chical RNN-based System for Modeling Expressive Piano Performance

    [Jeong et al., 2019] Dasaem Jeong, Taegyun Kwon, Yoojin Kim, Kyogu Lee, and Juhan Nam. VirtuosoNet: A Hierar- chical RNN-based System for Modeling Expressive Piano Performance. In Proceedings of the 20th International So- ciety for Music Information Retrieval Conference (ISMIR), Delft, Netherlands,

    work page 2019

  7. [16]

    V oicebox: Text-Guided Multilingual Uni- versal Speech Generation at Scale

    [Le et al., 2023] Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, and Wei Ning Hsu. V oicebox: Text-Guided Multilingual Uni- versal Speech Generation at Scale. In Advances in Neural Information Processing Systems,

    work page 2023

  8. [17]

    Content-based Controls for Music Large Lan- guage Modeling

    [Lin et al., 2024] Liwei Lin, Gus Xia, Junyan Jiang, and Yix- iao Zhang. Content-based Controls for Music Large Lan- guage Modeling. In Proceeding of the 25t International Society on Music Information Retrieval (ISMIR),

    work page 2024

Show all 40 references
  1. [18]

    Rendering music performance with interpretation variations using con- ditional variational RNN

    [Maezawa et al., 2019] Akira Maezawa, Kazuhiko Ya- mamoto, and Takuya Fujishima. Rendering music performance with interpretation variations using con- ditional variational RNN. Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR),

    2019

  2. [19]

    [Maman and Bermano, 2022] Ben Maman and Amit H. Bermano. Unaligned Supervision for Automatic Music Transcription in-the-Wild. In Proceedings of Machine Learning Research, volume 162,

    2022

  3. [20]

    Mustango: Toward Control- lable Text-to-Music Generation

    [Melechovsky et al., 2024] Jan Melechovsky, Zixun Guo, Deepanway Ghosal, Navonil Majumder, Dorien Herre- mans, and Soujanya Poria. Mustango: Toward Control- lable Text-to-Music Generation. In Proceedings of the 2024 Conference of the North American Chapter of the Association f...

    2024

  4. [21]

    Simulating Piano Performance Mistakes for Music Learning

    [Morsi et al., 2024] Alia Morsi, Huan Zhang, Akira Maezawa, Simon Dixon, and Xavier Serra. Simulating Piano Performance Mistakes for Music Learning. In Pro- ceedings of the Sound and Music Computing Conference (SMC),

    2024

  5. [22]

    Diff- A-Riff: Musical Accompaniment Co-creation via Latent Diffusion Models

    [Nistal et al., 2024] Javier Nistal, Marco Pasini, Cyran Aouameur, Maarten Grachten, and Stefan Lattner. Diff- A-Riff: Musical Accompaniment Co-creation via Latent Diffusion Models. Proceeding of the 25th International Society on Music Information Retrieval (ISMIR),

    2024

  6. [23]

    Anatomy of a perfor- mance: Sources of musical expression

    [Palmer, 1996] Caroline Palmer. Anatomy of a perfor- mance: Sources of musical expression. Music Perception, 13(3):433–453,

    1996

  7. [25]

    Automatic Note-Level Alignments in the ASAP Dataset

    [Peter et al., 2023] Silvan David Peter, Carlos Eduardo Cancino-chacón, Francesco Foscarin, Florian Henkel, and Gerhard Widmer. Automatic Note-Level Alignments in the ASAP Dataset. Transactions of the International Soci- ety for Music Information Retrieval (TISMIR),

    2023

  8. [26]

    Heinrich Neuhaus: A Life beyond Music

    [Razumovskaya, 2018] Maria Razumovskaya. Heinrich Neuhaus: A Life beyond Music . Boydell & Brewer, NED - new edition,

    2018

  9. [27]

    Differentiable Piano Model for MIDI-to- Audio Performance Synthesis

    [Renault et al., 2022] Lenny Renault, Rémi Mignot, and Axel Roebel. Differentiable Piano Model for MIDI-to- Audio Performance Synthesis. In Proceedings of the In- ternational Conference on Digital Audio Effects, DAFx , volume 3, pages 232–239,

    2022

  10. [28]

    Quantitative Effects of Global Tempo on Expressive Timing in Music Performance: Some Perceptual Evidence

    [Repp, 1995] Bruno H Repp. Quantitative Effects of Global Tempo on Expressive Timing in Music Performance: Some Perceptual Evidence. Music Perception, 13(1):39– 57,

    1995

  11. [30]

    Sketching the Expression: Flexible Rendering of Ex- pressive Piano Performance with Self-Supervised Learn- ing

    [Rhyu et al., 2022] Seungyeon Rhyu, Sarah Kim, and Kyogu Lee. Sketching the Expression: Flexible Rendering of Ex- pressive Piano Performance with Self-Supervised Learn- ing. In Proceeding of the International Society on Music Information Retrieval (ISMIR), Bengaluru, India,

    2022

  12. [31]

    GAPS: A Large and Diverse Classical Guitar Dataset and Benchmark Transcription Model

    [Riley et al., 2024] Xavier Riley, Zixun Guo, Drew Edwards, and Simon Dixon. GAPS: A Large and Diverse Classical Guitar Dataset and Benchmark Transcription Model. In Proceeding of the 25th International Society on Music In- formation Retrieval (ISMIR), aug

    2024

  13. [32]

    Joint Audio and Symbolic Conditioning for Temporally Controlled Text-to-Music Generation

    [Tal et al., 2024] Or Tal, Alon Ziv, Itai Gat, Felix Kreuk, and Yossi Adi. Joint Audio and Symbolic Conditioning for Temporally Controlled Text-to-Music Generation. In Pro- ceeding of the 25t International Society on Music Informa- tion Retrieval (ISMIR), jun

    2024

  14. [33]

    Audiobox: Unified Au- dio Generation with Natural Language Prompts

    [team, 2023] AudioBox team. Audiobox: Unified Au- dio Generation with Natural Language Prompts. Arxiv preprint arXiv:2312.15821,

    2023 arXiv

  15. [34]

    Continual learning of new sound classes using generative replay

    [Wang et al., 2019] Zhepei Wang, Cem Subakan, Efthymios Tzinis, Paris Smaragdis, and Laurent Charlin. Continual learning of new sound classes using generative replay. In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, volume 2019-Octob, jun

    2019

  16. [35]

    A Survey on Curriculum Learning

    [Wang et al., 2022] Xin Wang, Yudong Chen, and Wenwu Zhu. A Survey on Curriculum Learning. IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 44(9), oct

    2022

  17. [36]

    Large-Scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmenta- tion

    [Wu et al., 2023] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dub- nov. Large-Scale Contrastive Language-Audio Pretraining with Feature Fusion and Keyword-to-Caption Augmenta- tion. In ICASSP , IEEE International Conference on Acous- tics,...

    2023

  18. [37]

    [Wu et al., 2024] Shih Lun Wu, Chris Donahue, Shinji Watanabe, and Nicholas J. Bryan. Music ControlNet: Multiple Time-Varying Controls for Music Generation. IEEE/ACM Transactions on Audio Speech and Language Processing,

    2024

  19. [38]

    ATEPP: A Dataset of Automatically Transcribed Expressive Piano Performance

    [Zhang et al., 2022] Huan Zhang, Jingjing Tang, Syed Rafee, Simon Dixon, and George Fazekas. ATEPP: A Dataset of Automatically Transcribed Expressive Piano Performance. In Proceedings of the International Soci- ety for Music Information Retrieval Conference (ISMIR) , Bengaluru, India,

    2022

  20. [39]

    Adding Conditional Control to Text-to-Image Diffusion Models

    [Zhang et al., 2023] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In Proceedings of the IEEE Interna- tional Conference on Computer Vision,

    2023

  21. [40]

    LLaQo: Towards a Query-Based Coach in Expressive Music Per- formance Assessment

    [Zhang et al., 2025] Huan Zhang, Vincent Cheung, Hayato Nishioka, Simon Dixon, and Shinichi Furuya. LLaQo: Towards a Query-Based Coach in Expressive Music Per- formance Assessment. In In Proceedings of the IEEE In- ternational Conference on Acoustics, Speech and Signal Process...

    2025

  22. [1995]

    Pedal Timing and Tempo in Expressive Piano Performance: A Preliminary Investiga- tion

    [Repp, 1996] Bruno H Repp. Pedal Timing and Tempo in Expressive Piano Performance: A Preliminary Investiga- tion. Psychology of Music, 24(2):199–221,

    1996

  23. [1996]

    FiLM: Vi- sual reasoning with a general conditioning layer

    [Perez et al., 2018] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Vi- sual reasoning with a general conditioning layer. In 32nd AAAI Conference on Artificial Intelligence, AAAI 2018 , New Orleans, USA,

    2018

  24. [2002]

    On the Characterization of Expressive Per- formance in Classical Music: First Results of the Con Espressione Game

    [Cancino-Chacón et al., 2020] Carlos Cancino-Chacón, Sil- van Peter, Shreyan Chowdhury, Anna Aljanaki, and Ger- hard Widmer. On the Characterization of Expressive Per- formance in Classical Music: First Results of the Con Espressione Game. In Proceedings of the 21st Interna- t...

    2020

  25. [2008]

    Multi-instrument Music Syn- thesis with Spectrogram Diffusion

    [Hawthorne et al., 2022] Curtis Hawthorne, Ian Simon, Adam Roberts, Neil Zeghidour, Josh Gardner, Ethan Manilow, and Jesse Engel. Multi-instrument Music Syn- thesis with Spectrogram Diffusion. In Proceeding of the International Society on Music Information Retrieval (IS- MIR),...

    2022

  26. [2018]

    COCOLA: Coherence-Oriented Con- trastive Learning of Musical Audio Representations

    [Ciranni et al., 2025] Ruben Ciranni, Emilian Postolache, Giorgio Mariani, Michele Mancusi, Luca Cosmo, and Emanuele Rodolà. COCOLA: Coherence-Oriented Con- trastive Learning of Musical Audio Representations. In Proceeding of the IEEE International Conference on Acoustics, Spe...

    2025

  27. [2019]

    Deep Performer: Score-to-audio Music Performance Synthesis

    [Dong et al., 2022] Hao-Wen Dong, Cong Zhou, Taylor Berg-Kirkpatrick, and Julian Mcauley. Deep Performer: Score-to-audio Music Performance Synthesis. In Proceed- ing of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP),

    2022

  28. [2020]

    Computational Modeling of Expressive Music Perfor- mance with Linear and Non-linear Basis Function Models

    [Cancino-Chacón, 2018] Carlos Eduardo Cancino-Chacón. Computational Modeling of Expressive Music Perfor- mance with Linear and Non-linear Basis Function Models. PhD thesis, Johannes Kepler University Linz,

    2018

  29. [2022]

    Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons

    [Evans et al., 2024] Zach Evans, Julian D. Parker, CJ Carr, Zack Zukowski, Josiah Taylor, and Jordi Pons. Stable Au- dio Open. Arxiv preprint arXiv:2407.14358, jul

    2024 arXiv

  30. [2023]

    Characterizing Continual Learning Scenarios and Strategies for Audio Analysis

    [Bhatt et al., 2024] Ruchi Bhatt, Pratibha Kumari, Dwarikanath Mahapatra, Abdulmotaleb El Saddik, and Mukesh Saini. Characterizing Continual Learning Scenarios and Strategies for Audio Analysis. Arxiv preprint arXiv:2407.00465, jun

    2024 arXiv

  31. [2024]

    ScorePerformer : Expressive Piano Performance Render- ing with Fine-grained Control

    [Borovik and Viro, 2023] Ilya Borovik and Vladimir Viro. ScorePerformer : Expressive Piano Performance Render- ing with Fine-grained Control. In Proceeding of the 24th International Society on Music Information Retrieval (IS- MIR), Milan, Italy,

    2023

  32. [2025]

    Simple and Controllable Music Gen- eration

    [Copet et al., 2023] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez. Simple and Controllable Music Gen- eration. In Proceedings of the Conference on Neural Infor- mation Processing Systems (NeurIPS),

    2023

Pith tools

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