Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Recomposer: Event-roll-guided generative audio editing

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read One transformer deletes, inserts, and enhances single sounds in a dense scene using text plus a time-aligned event roll.

desk verdict Solid, honest systems paper; the activity-roll conditioning is genuinely new and it works on synthetic scenes, but real-scene editing is untested and the paper knows it. read the letter →

arxiv 2509.05256 v1 pith:2YOQGCEI submitted 2025-09-05 cs.SD cs.AIcs.LGeess.AS

Daniel P. W. Ellis , Eduardo Fonseca , Ron J. Weiss , Kevin Wilson , Scott Wisdom , Hakan Erdogan , John R. Hershey , Aren Jansen
show 2 more authors
R. Channing Moore Manoj Plakal
This is my paper
classification cs.SDcs.AIcs.LGeess.AS
keywords audioeditinggenerativesoundeventactivityrollStreamtransformerdeleteenhanceinsertscenerecomposition
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 tries to establish that editing sound scenes can be reframed as "recomposition": instead of cutting and pasting waveforms, a generative model reads the original audio, a short text instruction (e.g., "delete dog bark"), and a time-aligned event roll, then synthesizes a new waveform that applies the requested edit. The authors train an encoder-decoder transformer on tens of thousands of synthetic mixtures formed by adding isolated Freesound foreground events to dense AudioSet backgrounds, with Delete, Insert, and Enhance conditions. The central claim is that this synthetic training is enough for the model to remove, insert, and boost individual events, while leaving untouched regions nearly unmodified, and that the action, class, and timing information each measurably contribute. If true, it gives a practical interface for editing real-world recordings at the level of perceived sound events rather than raw audio.

What carries the argument

The central mechanism is an encoder-decoder transformer operating on SoundStream residual-vector-quantized (RVQ) audio tokens. The input audio is encoded into a time-aligned stack of codec features, and the edit instructions are embedded with a frozen Sentence-T5 text encoder, then projected onto the time axis via a binary activity roll that marks which frames are edited. An autoregressive RQ-transformer decoder generates the output tokens frame by frame, with a depth transformer producing each residual token conditioned on the previous ones, and the SoundStream decoder turns those tokens back into a waveform. The activity roll is what carries the precise temporal control: it is a vector of

What would settle it

Take a real-world recording with an overlapping event (e.g., a dog bark during speech), run the model with a ground-truth event roll, and listen for whether the edit is applied cleanly without artifacts in nontarget regions; if the edited output noticeably distorts the background or fails to remove the event, the synthetic-mixture assumption is falsified. A quantitative version would compare Delete/Insert performance on a diverse set of real recordings (with human-labeled event times) to performance on synthetic mixtures with matched class and TBR.

Watch

Extended reading notes

Core claim

A generative audio model can modify individual sound events in a complex scene when conditioned on an edit description that pairs free-text instructions with explicit time extents. The model is trained on synthetic input/desired-output pairs built by linearly mixing isolated target events into dense real-world backgrounds. In the target time regions, the model's estimates improve over the unprocessed input on both a waveform-level metric and a classifier-based divergence metric, for deletion, insertion, and enhancement. The time-aligned activity roll is what provides millisecond-level control, and ablation experiments show that timing is the most critical conditioning signal, action matters

Load-bearing premise

The load-bearing premise is that linearly mixing isolated Freesound events into dense AudioSet backgrounds at hand-set target-to-background ratios produces training examples representative enough of real-world sound scenes and real editing effects.

Editorial extensions

If this is right

  • If the central claim holds, a single model can perform delete, insert, and enhance edits without separate systems for separation, generation, and mixing, because the activity roll resolves which event to affect.
  • Time-aligned event control at the frame level becomes possible with free-text action/class descriptions, going beyond text-only timing cues like "at the beginning" used by earlier audio editors.
  • The synthetic training recipe—isolated foreground events mixed into dense backgrounds at controlled target-to-background ratios—is sufficient to teach deletion, insertion, and enhancement behavior, including generalization to held-out backgrounds and target events.
  • Ablation results imply that timing is the most load-bearing conditioning signal; without it the model cannot reliably target the right event, while removing class identity harms insertion more than deletion or enhancement.
  • The system's limitation to a fixed vocabulary of AudioSet class labels and a single output level (15 dB target-to-background ratio) is a recognized gap; overcoming it requires richer training descriptions and explicit control over generated event properties.

Reading between the lines

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

  • If additive mixing is a reasonable approximation for real scenes, this approach could evolve into a full recomposition editor where event rolls come from an automatic recognizer rather than ground truth; the paper leaves that recognition-to-roll path untested, so the practical gain rests on how much recognizer errors degrade edit fidelity.
  • The decoy-event evaluation suggests a natural stress test: scenes with two similar or same-class events at different times would reveal whether timing truly disambiguates identity or whether the model relies on acoustic prominence differences.
  • A concrete extension is to make the output level controllable by conditioning on a desired target-to-background ratio, turning the current fixed 15 dB output into a user-settable parameter.
  • The transition from Enhance at very low input TBR (essentially text-to-audio generation from the class name) to Enhance at higher TBR (closer to source separation) suggests a continuum that could be exploited by conditioning on estimated target level, a knob the current model does not expose.
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 presents Recomposer, an encoder-decoder transformer operating on SoundStream audio tokens, conditioned on an 'activity roll' that specifies edit actions (delete, insert, enhance), target event class labels, and precise time extents through text embeddings from Sentence-T5. Training pairs are generated synthetically by mixing isolated Freesound target events into dense AudioSet backgrounds at controlled target-to-background ratios. The model is evaluated on frozen synthetic test sets with multiscale spectral distortion (MSD) and YAMNet-based KL divergence, reporting consistent target-region improvements over the unprocessed input for all three actions. Ablation studies with decoy events show the contributions of timing, action, and class conditioning. The authors also report informal listening on non-synthetic scenes and explicitly describe the system as a proof-of-concept.

Significance. If the demonstrated behavior transfers beyond the synthetic setup, the paper makes a useful contribution: it shows that a single generative model can perform three distinct event-level edits with precise temporal control, and its decoy-based ablation methodology is a thoughtful way to isolate the role of conditioning information. Strengths include the on-the-fly synthetic data generation at scale, the use of frozen pretrained codec and text encoders, the clear reporting of target versus nontarget regions, and the public release of audio examples. The main significance is limited by the fact that all quantitative results are obtained on synthetic mixtures with deliberately mismatched recording conditions; this leaves open the possibility that the model exploits low-level acoustic mismatch rather than semantic event representations. The paper is honest about this limitation, but the central claim of editing 'complex real-world sound scenes' is not yet quantitatively supported.

major comments (3)
  1. [§3.1, Table 1] The central quantitative evidence is obtained exclusively on synthetic mixtures where target events (Freesound) and backgrounds (AudioSet) come from unrelated recording conditions. Section 3.1 explicitly concedes that 'the unrelated recording conditions of target and background could give unnatural cues to help the model identify the target portion.' This is a serious confound: the strong target-region gains (e.g., Delete MSD 4.8→2.5, KLD 1.6→0.5; Enhance MSD 3.4→2.6, KLD 1.6→0.9) could reflect detection of reverb, noise-floor, or channel mismatches rather than representation of the event itself. To support the paper's claim of editing real-world sound scenes, the authors should either add a controlled evaluation on real recordings with human-annotated events or construct a matched-condition synthetic set (e.g., mixing targets from the same recording or channel-matched data) to show the
  2. [§3.1, §4 (KLD metric)] YAMNet is used both to curate target events (Figure 3 red box, via classifier score thresholding) and to compute the KLD evaluation metric. This creates a circularity: training targets are selected to be recognizable by YAMNet, and the model is rewarded for generating outputs that the same classifier recognizes. This can inflate the reported class-similarity improvements, especially for Insert and Enhance. The authors should evaluate with an independent sound-event classifier not used in data curation, or at minimum report results with a held-out classifier to quantify the bias. The paper currently does not acknowledge this issue.
  3. [§4.1, Table 1, Figure 5] All reported metrics are point estimates with no confidence intervals, significance tests, or variance estimates. Given that each evaluation set has 10,000 examples, bootstrap confidence intervals are easy to compute. This is particularly important for the ablation study in Figure 5, where differences between conditions (e.g., whether class information helps Delete and Enhance) appear small; without uncertainty quantification, the claim that 'the results are broadly in line with expectations' is not rigorously supported. This is a load-bearing issue for the ablation conclusions.
minor comments (6)
  1. [Section 4] "Kullback-Liebler" should be "Kullback-Leibler".
  2. [Figure 2 and Section 3] The notation (d_i, t) for the instruction matrix is not fully defined. State explicitly that d_i is the Sentence-T5 embedding dimension and t is the number of time frames in the SoundStream encoding.
  3. [Table 1] The nontarget input rows are 0.0 by construction. This is clear but could be stated in the caption to avoid confusion.
  4. [Figure 5] The x-axis labels 'T / A / C', 'T / A / -', etc. are terse. Adding a legend or a table mapping abbreviations to conditioning components would improve readability.
  5. [Abstract, Section 5] The abstract calls the work 'an important and practical application', while Section 5 says it is 'more of a proof-of-concept than a practical tool'. Please align the language with the stated scope.
  6. [References] Reference [28] is a URL without a year or formal citation; consider citing a published description or technical report.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical evaluations of a trained transformer, with no derivation-level reduction to its inputs.

full rationale

Recomposer makes no first-principles derivation; its claims are empirical evaluations of a trained transformer. The training/evaluation pairs are synthetic mixtures, and the paper explicitly concedes in Section 3.1 that 'the unrelated recording conditions of target and background could give unnatural cues' and in Section 5 that the system is 'more of a proof-of-concept than a practical tool.' These are generalization and validity risks, not circularity. The only potential coupling is that YAMNet is used both to select target events in Section 3.1 and as the KLD evaluation classifier in Section 4; however, the model is trained to match desired waveforms via token prediction, not to optimize YAMNet KLD, so the reported KLD improvements are not forced by construction. Self-citations to AudioSet, YAMNet, and FSD50K are citations to public, externally usable datasets and models and do not carry the load of an unverified uniqueness claim. No equation-level reduction, fitted-parameter-renamed-as-prediction, or ansatz-smuggled-via-self-citation is present.

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

The central claim rests on hand-set data-generation choices, the fidelity of frozen codec and text encoders, and the assumption that synthetic additive mixtures stand in for real edits. No equation-level free parameters are present because the work is an empirical system paper, and no new physical entities are introduced.

free parameters (4)
  • Training target-to-background ratios (TBR) = Delete input 10 dB; Enhance input -6 dB (or uniform -30 to 0 dB in the enhancement-only model); desired output 10 or 15
    Hand-set in Section 3.1 to make edit tasks visible; directly controls task difficulty and defines what "enhance" means.
  • Target event selection thresholds = Event duration 0.2-2.0 s; energy envelope above noise floor; hysteresis thresholding on classifier scores; backgrounds w
    Curation rules in Section 3.1 determine the target pool and therefore which event classes and durations are learnable.
  • Number and composition of edits per training example = Two independent edits uniformly chosen from {enhance, delete, insert, no-op} (6.25% no-op pairs)
    Section 3.1; design choice that sets multi-edit capability and the no-op rate.
  • Desired output level for enhancement = 15 dB TBR
    Section 4 and Figure 4; model always outputs a high-level target, limiting practical control over generated event properties, acknowledged in Section 5.
assumptions (4)
  • domain assumption Synthetic linear mixtures of Freesound targets and AudioSet backgrounds are representative of real sound scenes and editing operations.
    All training and evaluation pairs are generated this way (Section 3.1); Section 5 admits proof-of-concept status, so transfer to real scenes is unverified.
  • domain assumption The event roll is provided as ground truth by an external recognizer.
    Section 1 defers sound event recognition to prior systems; Section 5 confirms interactive use relied on ground-truth transcripts. No noisy-roll experiments are reported.
  • domain assumption SoundStream encodes enough information for the task and its reconstruction distortion is perceptually minimal.
    The model operates and is evaluated on SoundStream codes (Sections 3 and 4); if the codec removes event detail, the model cannot perform the edits.
  • domain assumption YAMNet classifier posteriors are a valid measure of class-level edit success.
    The KLD metric uses YAMNet (Section 4), and YAMNet was also used to select target events in training data (Section 3.1), creating shared-classifier bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recomposer: Event-roll-guided generative audio editing." pith.science (2026). https://pith.science/paper/2YOQGCEI

@misc{pith2026250905256,
  author       = {Pith},
  title        = {Pith review of: Recomposer: Event-roll-guided generative audio editing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2YOQGCEI}},
  note         = {Machine review of arXiv:2509.05256}
}
read the original abstract

Editing complex real-world sound scenes is difficult because individual sound sources overlap in time. Generative models can fill-in missing or corrupted details based on their strong prior understanding of the data domain. We present a system for editing individual sound events within complex scenes able to delete, insert, and enhance individual sound events based on textual edit descriptions (e.g., ``enhance Door'') and a graphical representation of the event timing derived from an ``event roll'' transcription. We present an encoder-decoder transformer working on SoundStream representations, trained on synthetic (input, desired output) audio example pairs formed by adding isolated sound events to dense, real-world backgrounds. Evaluation reveals the importance of each part of the edit descriptions -- action, class, timing. Our work demonstrates ``recomposition'' is an important and practical application.

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. MMAE: A Massive Multitask Audio Editing Benchmark

    cs.SD 2026-06 conditional novelty 8.0 of 10

    MMAE is a new multitask audio editing benchmark showing that leading models achieve under 5% exact match rate, with 0% on complex mixed-modality tasks.

  2. RIME: Enabling Large-Scale Agentic Music Post-Production

    cs.SD 2026-07 conditional novelty 6.0 of 10

    RIME generates 3,000 synthetic music post-production edit triples and shows that current multimodal LLM agents can recover edit structure but often fail to set effect parameters correctly.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages · cited by 2 Pith papers

  1. [1]

    Listen, think, and understand,

    Y . Gong, H. Luo, A. H. Liu, L. Karlinsky, and J. Glass, “Listen, think, and understand,” inProc. ICLR, 2024

    work page 2024

  2. [2]

    Text-driven separation of arbitrary sounds,

    K. Kilgour, B. Gfeller, Q. Huang, A. Jansen, S. Wisdom, and M. Tagliasac- chi, “Text-driven separation of arbitrary sounds,” inProc. Interspeech, 2022

    work page 2022

  3. [3]

    AudioGen: Textually guided audio generation,

    F. Kreuk, G. Synnaeve, A. Polyak, U. Singer, A. D ´efossez, J. Copet, D. Parikh, Y . Taigman, and Y . Adi, “AudioGen: Textually guided audio generation,” inProc. ICLR, 2023

    work page 2023

  4. [4]

    AudioLDM: Text-to-audio generation with latent diffusion models,

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “AudioLDM: Text-to-audio generation with latent diffusion models,” inProc. ICML, 2023, pp. 21 450–21 474

    work page 2023

  5. [5]

    Text-to-audio generation using instruction guided latent diffusion model,

    D. Ghosal, N. Majumder, A. Mehrish, and S. Poria, “Text-to-audio generation using instruction guided latent diffusion model,” inProc. ACM Multimedia, 2023, pp. 3590–3598

    work page 2023

  6. [6]

    AudioLDM 2: Learning holistic audio generation with self-supervised pretraining,

    H. Liu, Y . Yuan, X. Liu, X. Mei, Q. Kong, Q. Tian, Y . Wang, W. Wang, Y . Wang, and M. D. Plumbley, “AudioLDM 2: Learning holistic audio generation with self-supervised pretraining,”IEEE/ACM Trans. Audio, Speech, Lang. Process., 2024

    work page 2024

  7. [7]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” inProc. CVPR, 2023, pp. 3836–3847

    work page 2023

  8. [8]

    Uni-ControlNet: All-in-one control to text-to-image diffusion models,

    S. Zhao, D. Chen, Y .-C. Chen, J. Bao, S. Hao, L. Yuan, and K.-Y . K. Wong, “Uni-ControlNet: All-in-one control to text-to-image diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 11 127–11 150, 2023

    work page 2023

Show all 35 references
  1. [9]

    Emu Edit: Precise image editing via recognition and generation tasks,

    S. Sheynin, A. Polyak, U. Singer, Y . Kirstain, A. Zohar, O. Ashual, D. Parikh, and Y . Taigman, “Emu Edit: Precise image editing via recognition and generation tasks,” inProc. CVPR, 2024, pp. 8871–8879

    2024

  2. [10]

    AUDIT: Audio editing by following instructions with latent diffusion models,

    Y . Wang, Z. Ju, X. Tan, L. He, Z. Wu, J. Bian, and S. Zhao, “AUDIT: Audio editing by following instructions with latent diffusion models,” in Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 71 340–71 357

    2023

  3. [11]

    Music ControlNet: Multiple time-varying controls for music generation,

    S.-L. Wu, C. Donahue, S. Watanabe, and N. J. Bryan, “Music ControlNet: Multiple time-varying controls for music generation,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 32, pp. 2692–2703, 2024

    2024

  4. [12]

    Instruct- MusicGen: Unlocking text-to-music editing for music language models via instruction tuning,

    Y . Zhang, Y . Ikemiya, W. Choi, N. Murata, M. A. Mart ´ınez-Ram´ırez, L. Lin, G. Xia, W.-H. Liao, Y . Mitsufuji, and S. Dixon, “Instruct- MusicGen: Unlocking text-to-music editing for music language models via instruction tuning,”arXiv preprint arXiv:2405.18386, 2024

    2024 arXiv

  5. [13]

    Sketch2Sound: Controllable audio generation via time-varying signals and sonic imitations,

    H. F. Garc ´ıa, O. Nieto, J. Salamon, B. Pardo, and P. Seetharaman, “Sketch2Sound: Controllable audio generation via time-varying signals and sonic imitations,”arXiv preprint arXiv:2412.08550, 2024

    2024 arXiv

  6. [14]

    Audio- Composer: Towards fine-grained audio generation with natural language descriptions,

    Y . Wang, H. Chen, D. Yang, Z. Wu, H. Meng, and X. Wu, “Audio- Composer: Towards fine-grained audio generation with natural language descriptions,” inProc. ICASSP, 2025

    2025

  7. [15]

    PicoAudio: Enabling precise timestamp and frequency controllability of audio events in text-to-audio generation,

    Z. Xie, X. Xu, Z. Wu, and M. Wu, “PicoAudio: Enabling precise timestamp and frequency controllability of audio events in text-to-audio generation,”arXiv preprint arXiv:2407.02869, 2024

    2024 arXiv

  8. [16]

    AudioLM: a language modeling approach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchiet al., “AudioLM: a language modeling approach to audio generation,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 2523–2533, 2023

    2023

  9. [17]

    SoundStream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “SoundStream: An end-to-end neural audio codec,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 30, pp. 495–507, 2021

    2021

  10. [18]

    Neural machine translation by jointly learning to align and translate,

    D. Bahdanau, K. Cho, and Y . Bengio, “Neural machine translation by jointly learning to align and translate,” inProc. ICLR, 2015

    2015

  11. [19]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  12. [20]

    Sentence-T5: Scalable sentence encoders from pre-trained text-to-text models,

    J. Ni, G. Hernandez Abrego, N. Constant, J. Ma, K. Hall, D. Cer, and Y . Yang, “Sentence-T5: Scalable sentence encoders from pre-trained text-to-text models,” inProc. ACL, May 2022

    2022

  13. [21]

    Autoregressive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregressive image generation using residual quantization,” inProc. CVPR, 2022

    2022

  14. [22]

    Moshi: a speech-text foundation model for real-time dialogue,

    A. D´efossez, L. Mazar´e, M. Orsini, A. Royer, P. P´erez, H. J´egou, E. Grave, and N. Zeghidour, “Moshi: a speech-text foundation model for real-time dialogue,”arXiv preprint arXiv:2410.00037, 2024

    2024 arXiv

  15. [23]

    Freesound technical demo,

    F. Font, G. Roma, and X. Serra, “Freesound technical demo,” inProc. ACM Multimedia, 2013, pp. 411–412

    2013

  16. [24]

    Unsupervised sound separation using mixture invariant training,

    S. Wisdom, E. Tzinis, H. Erdogan, R. J. Weiss, K. Wilson, and J. Hershey, “Unsupervised sound separation using mixture invariant training,” in Advances in Neural Information Processing Systems, vol. 33, 2020, pp. 3846–3857

    2020

  17. [25]

    Evaluation of algorithms using games: the case of music annotation,

    E. Law, K. West, M. Mandel, M. Bay, and J. Downie, “Evaluation of algorithms using games: the case of music annotation,” inProc. ISMIR, 2010

    2010

  18. [26]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProc. NAACL, 2019

    2019

  19. [27]

    The benefit of temporally-strong labels in audio event classification,

    S. Hershey, D. P. W. Ellis, E. Fonseca, A. Jansen, C. Liu, R. Chan- ning Moore, and M. Plakal, “The benefit of temporally-strong labels in audio event classification,” inProc. ICASSP, 2021, pp. 366–370

    2021

  20. [28]

    Sound classification with Y AMNet,

    M. Plakal and D. P. W. Ellis, “Sound classification with Y AMNet,” 2020. [Online]. Available: https://www.tensorflow.org/hub/tutorials/yamnet

    2020

  21. [29]

    FSD50K: an open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “FSD50K: an open dataset of human-labeled sound events,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 30, pp. 829–852, 2022

    2022

  22. [30]

    Neural source-filter waveform models for statistical parametric speech synthesis,

    X. Wang, S. Takaki, and J. Yamagishi, “Neural source-filter waveform models for statistical parametric speech synthesis,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 28, pp. 402–415, 2019

    2019

  23. [31]

    DDSP: Differentiable digital signal processing,

    J. Engel, L. Hantrakul, C. Gu, and A. Roberts, “DDSP: Differentiable digital signal processing,” inProc. ICLR, 2020

    2020

  24. [32]

    Diffsound: Discrete diffusion model for text-to-sound generation,

    D. Yang, J. Yu, H. Wang, W. Wang, C. Weng, Y . Zou, and D. Yu, “Diffsound: Discrete diffusion model for text-to-sound generation,” IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 31, pp. 1720– 1733, 2023

    2023

  25. [33]

    Fr ´echet audio distance: A metric for evaluating music enhancement algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr ´echet audio distance: A metric for evaluating music enhancement algorithms,” in Proc. Interspeech, 2019

    2019

  26. [34]

    Adapting Fr ´echet audio distance for generative music evaluation,

    A. Gui, H. Gamper, S. Braun, and D. Emmanouilidou, “Adapting Fr ´echet audio distance for generative music evaluation,” inProc. ICASSP, 2024, pp. 1331–1335

    2024

  27. [35]

    Correlation of Fr ´echet audio distance with human perception of environmental audio is embedding dependent,

    M. Tailleur, J. Lee, M. Lagrange, K. Choi, L. M. Heller, K. Imoto, and Y . Okamoto, “Correlation of Fr ´echet audio distance with human perception of environmental audio is embedding dependent,” inProc. EUSIPCO, 2024, pp. 56–60

    2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.