Pith. sign in

REVIEW 4 major objections 5 minor 40 references

DGMO: Training-Free Audio Source Separation through Diffusion-Guided Mask Optimization

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A pretrained audio diffusion model can separate sources from a text query with no training beyond its original generation task.

desk verdict Novel zero-shot LASS via a frozen diffusion model, but the reference-driven mask optimization needs per-sample validation before the 'inherent separation' claim holds. read the letter →

arxiv 2506.02858 v2 pith:JC464XJY submitted 2025-06-03 eess.AS cs.AIcs.SD

Geonyoung Lee , Geonhee Han , Paul Hongsuck Seo This is my paper
classification eess.AScs.AIcs.SD
keywords language-queriedaudiosourceseparationzero-shotdiffusionmodeltest-timeoptimizationspectrogrammaskDDIMinversiontraining-freetext-to-audio
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 text-to-audio diffusion model, trained only to generate sound, can also be used at test time to pull a single described source out of a mixed audio clip, with no task-specific training. The proposed method, DGMO, first generates several reference renditions of the queried source from the mixture, then optimizes a spectrogram mask so that masking the mixture reproduces those references. Because the mask acts on the mixture's magnitude spectrogram and reconstruction keeps the mixture's original phase, the result stays aligned with the input while the diffusion prior supplies semantic guidance. On VGGSound, AudioCaps, MUSIC, and ESC-50 mixtures, this training-free approach reports separation quality competitive with some supervised LASS systems and clearly better than naive diffusion-based adaptations. If the claim holds, open-vocabulary audio separation becomes a zero-shot capability of generative models rather than a task requiring dedicated labeled data.

What carries the argument

The central object is the learnable mask $M$ applied to the mixture's magnitude spectrogram, optimized by gradient descent against a mel-domain reconstruction loss using diffusion-generated references. Two design choices carry the argument: decoupling the optimization space (magnitude spectrogram, which is invertible and keeps the mixture's phase) from the supervision space (mel spectrogram, where the diffusion model operates), and using DDIM inversion instead of random noise injection to produce references that preserve the mixture's content structure. Averaging the loss over multiple references reduces the variance of the mask optimization, since individual references capture different aspects of the target source.

What would settle it

On a synthetic mixture of two known sounds (for example, a cat meow and a dog bark mixed at 0 dB SNR), query the cat, run DGMO, and compare the output to the ground-truth cat and to the oracle mask applied to the mixture: if the method is right, the separated waveform should approach the oracle-masked mixture, whereas if the diffusion reference hallucinates, the output will contain spectral energy in time-frequency bins where neither source has energy and the SI-SDR gain will drop for queries where hallucination is systematic.

Watch

Extended reading notes

Core claim

The central discovery is that pretrained diffusion models for text-to-audio generation already encode enough text-to-sound knowledge to perform language-queried source separation without any separation-specific training. DGMO makes this concrete in two stages: a reference generation stage uses DDIM inversion to noise the mixture deterministically and then denoises it under the text query, producing reference audio that inherits the mixture's structure while being regenerated toward the target source; a mask optimization stage then learns a mask $M$ over the mixture's magnitude spectrogram by minimizing $\mathcal{L}_i(M) = \| \operatorname{mel}(x_{\mathrm{spec}} \odot M) - s_i^{\mathrm{mel}} \|_2^2$ against the mel-spectrograms of one or more references. The final separated waveform is reconstructed as $\hat{s}^* = \operatorname{iSTFT}(x_{\mathrm{phase}}, x_{\mathrm{spec}} \odot M^*)$, retaining the mixture's true phase. The paper reports that this procedure lifts SI-SDR from near zero to roughly 2 dB on AudioCaps and works across three different audio diffusion backbones, with separation quality tracking generation quality. The authors also acknowledge that references often contain sounds not present in the original mixture, and the mask optimization is precisely the mechanism that tries to suppress such hallucinated content.

Load-bearing premise

The method assumes the diffusion-generated reference signal is a valid separation target, meaning its mel spectrogram corresponds to a sound actually present in the mixture; the paper itself notes that references often introduce sound elements not originally in the mixture, and if that happens systematically the mask optimization will inject those absent sounds into the output.

Editorial extensions

If this is right

  • A frozen text-to-audio diffusion model can serve as a zero-shot LASS system without separation-specific training data or fine-tuning.
  • Separation quality tracks the generation quality of the underlying diffusion model, so improvements in text-to-audio generation should directly translate into better zero-shot separation.
  • Retaining the mixture's original phase during reconstruction avoids the temporal artifacts and phase errors that arise when separated audio is generated directly or reconstructed from a mel spectrogram.
  • Using multiple diffusion-generated references as supervision stabilizes mask optimization, making the method more robust to the variance of individual generations.

Reading between the lines

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

  • The same two-space decoupling — optimizing an invertible mask while supervising in a model's native compressed domain — could transfer to other inverse problems with generative priors, such as speech enhancement, audio restoration, or music source separation with text queries.
  • A testable boundary of the zero-shot claim is that the method inherits the diffusion prior's biases: on sound classes underrepresented in text-to-audio training data, reference generation may hallucinate, and the mask optimizer would then inject those hallucinated sounds into the output even when the true source is present in the mixture.
  • The paper's results suggest a compute-versus-quality trade-off that could be pushed further: increasing the number of references and optimization steps may approach oracle masking on simple mixtures, and the residual gap would then measure reference hallucination rather than mask optimizer capacity.
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

4 major / 5 minor

Summary. The paper proposes DGMO, a training-free framework for language-queried audio source separation that combines a frozen pretrained text-to-audio diffusion model with test-time mask optimization. A query-conditioned reference signal is generated from the mixture via DDIM inversion and denoising, and a mask in the magnitude spectrogram domain is optimized so that the mel spectrogram of the masked mixture matches the reference. The method is evaluated on VGGSound, AudioCaps, MUSIC, and ESC-50, with comparisons to naive baselines and supervised systems, and ablations over diffusion backbones and noising-step ratios.

Significance. If the central claim holds, this is a useful demonstration that a frozen generative prior can serve as a zero-shot separation guide, and the paper provides a concrete recipe with code. Strengths include the use of publicly available components, evaluation on four datasets, and ablations showing robustness across diffusion models. The significance is conditional, however, because the load-bearing assumption that the diffusion-generated reference is a valid separation target is not directly verified, and the quantitative evidence is weakened by single-mean reporting and hyperparameter selection on a small subset.

major comments (4)
  1. [Section 3.4, Eq. (4)] The optimization objective treats the mel spectrogram of the diffusion-generated reference, smel_i, as a ground-truth target for the mask. The authors themselves state that references 'often introduce sound elements that are not originally present in x', yet no analysis shows that the reference is closer in time-frequency structure to the true source than to the mixture or to non-target sources. Without such a check, the reported SI-SDR/SDRi gains are also consistent with the mask selecting query-relevant but source-inaccurate components. Please add a reference-fidelity study, for example mel-distance between the reference and the true target versus the mixture, per-sample failure analysis, or an oracle-mask upper bound, to support the claim that the model 'inherently' separates rather than merely selects semantically related content.
  2. [Section 4.3 and Table 4] The noising-step ratio t/T = 0.7 is fixed in Section 4.3, but Table 4 reports results on only 100 AudioCaps samples and no selection rule or validation split is given. Notably, for DDIM inversion the table shows t/T = 0.9 yielding slightly higher SI-SDR (2.04 vs. 1.99) and SDRi (3.64 vs. 3.57), so the basis for choosing 0.7 is unclear. The mask optimization epochs (300) and reference batch size (4) are also not ablated. Please report how these hyperparameters were selected, provide per-dataset sensitivity, or otherwise demonstrate that the reported configuration is not tuned to the evaluation subset.
  3. [Tables 1-4] All results are reported as single means without variance, confidence intervals, or significance tests. Since some comparisons involve small differences, for example Table 3 shows SDRi values of 3.12, 2.89, and 3.57 across diffusion backbones, it is difficult to assess whether the improvements are robust. Please report standard deviations, per-sample distributions, or significance tests at least for the main comparisons in Tables 1 and 2.
  4. [Table 2 and Section 4.4] The supervised baseline numbers for LASS-Net, CLIPSep, and AudioSep are 'sourced from [5, 31]' and CLAP scores from [5], but the paper does not establish that the mixture construction, evaluation code, and metric versions are identical to those used for DGMO. If the protocols differ, the 'competitive performance' claim is not supported. Please recompute baselines under the identical protocol or document the exact provenance and confirm that the evaluation conditions match.
minor comments (5)
  1. [Section 3.2, Eq. (1)] The symbols t and T are used without explicit definitions; please define t as the diffusion timestep and T as the total number of steps, and clarify the relation to the noising-step ratio t/T used later.
  2. [Section 4.3] The phrase 'with a noising step ratio of 0.7 and null text' is ambiguous: please clarify whether null text is used during DDIM inversion and how the language query is injected during the subsequent denoising reference generation.
  3. [Tables 1 and 3] Several entries contain formatting artifacts, such as '−0.06 2 .33' and '1.10 3 .12'; please fix the spacing and alignment in the tables.
  4. [References] Reference [31] is listed only as 'CoRR' without volume, article number, or year; please provide a complete citation.
  5. [Table 3] The FAD values are taken from [8], but the dataset and checkpoint used for those FAD computations are not specified, which makes the claimed correlation between generation quality and separation performance difficult to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the mask is optimized against a diffusion-generated reference, but evaluation uses external ground-truth sources, so the central claim is independently tested.

full rationale

The derivation chain is self-contained. Reference generation (Section 3.3) produces a query-guided, DDIM-inverted denoised version of the mixture. Mask optimization (Section 3.4, Eqs. 4-5) fits a learnable magnitude mask so that the mel spectrogram of the masked mixture matches the reference, and reconstruction (Eq. 6) retains the mixture's original phase. No parameter is fitted to the ground-truth separation targets during development, and all reported numbers are computed against held-out true sources using SI-SDR, SDRi, and CLAP Score, which are external to the optimization objective. The only self-referential element is that the same pretrained diffusion model generates the supervision reference, but this is the method under test rather than a fitted input that predetermines the evaluation outcome. The paper explicitly acknowledges in Section 3.4 that references 'often introduce sound elements that are not originally present in x'; this is an honest correctness risk about reference fidelity, not a circularity, because the final metrics would penalize such hallucinated content. There are no load-bearing self-citations, no imported uniqueness theorems, and no equations that reduce to their own inputs by construction. The central claim that a frozen text-to-audio diffusion model can guide training-free separation is therefore empirically falsifiable rather than circular.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new mathematical constants or physical entities. Its free parameters are test-time optimization hyperparameters, none derived from first principles. The load-bearing assumptions are about the transferability of the generative prior, the validity of phase-preserving masking, the sufficiency of the mel-distance objective, and the behavior of DDIM inversion.

free parameters (3)
  • Noising step ratio (t/T) = 0.7
    Selected on a 100-sample AudioCaps test subset in Table 4; values 0.5 and 0.9 give SDRi 3.15 and 3.64, so 0.7 is hand-tuned on test data rather than derived.
  • Mask optimization epochs = 300 per iteration, 2 iterations
    Reported in Section 4.3 with no sensitivity analysis.
  • Reference batch size = 4
    Reported in Section 4.3; no ablation of the number of references.
assumptions (5)
  • domain assumption A pretrained text-to-audio diffusion model carries a generative prior that can localize a queried source inside a mixture
    This is the premise of the whole framework, stated in Sections 1 and 3.4; if the prior cannot separate source semantics, references will be useless.
  • domain assumption Magnitude-spectrogram masking with the mixture's original phase reconstructs a faithful separated waveform
    Used in Eq. 6 and the mask application; standard in LASS but phase from the mixture can leak or distort the target.
  • ad hoc to paper Mel-domain distance between the masked mixture and the reference is a sufficient objective for separation
    Eq. 4 defines the entire optimization objective; the paper gives no independent justification beyond the experiments.
  • domain assumption DDIM inversion with null text preserves input content while allowing query-conditioned regeneration
    Section 3.4 relies on this to generate references; the ablation in Table 4 supports it empirically but it remains an assumption about the inversion procedure.
  • standard math STFT, mel filtering, and iSTFT compose into a differentiable pipeline with acceptable loss
    Standard signal processing assumptions; mel is lossy and non-invertible, which is why the paper decouples optimization and evaluation domains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DGMO: Training-Free Audio Source Separation through Diffusion-Guided Mask Optimization." pith.science (2026). https://pith.science/paper/JC464XJY

@misc{pith2026250602858,
  author       = {Pith},
  title        = {Pith review of: DGMO: Training-Free Audio Source Separation through Diffusion-Guided Mask Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JC464XJY}},
  note         = {Machine review of arXiv:2506.02858}
}
read the original abstract

Language-queried Audio Source Separation (LASS) enables open-vocabulary sound separation via natural language queries. While existing methods rely on task-specific training, we explore whether pretrained diffusion models, originally designed for audio generation, can inherently perform separation without further training. In this study, we introduce a training-free framework leveraging generative priors for zero-shot LASS. Analyzing naive adaptations, we identify key limitations arising from modality-specific challenges. To address these issues, we propose Diffusion-Guided Mask Optimization (DGMO), a test-time optimization framework that refines spectrogram masks for precise, input-aligned separation. Our approach effectively repurposes pretrained diffusion models for source separation, achieving competitive performance without task-specific supervision. This work expands the application of diffusion models beyond generation, establishing a new paradigm for zero-shot audio separation. The code is available at: https://wltschmrz.github.io/DGMO/

Figures

Figures reproduced from arXiv: 2506.02858 by the authors.

Figure 1
Figure 1. Training-free LASS framework using pre-trained diffusion model. It has two key processes: a Reference Generation and a Mask Optimization. 3. Method 3.1. Language-queried Audio Source Separation Given an audio mixture x composed of multiple source signals {si} and environmental noise e formulated as x = P i si + e, LASS [2] aims to extract a target source s ∗ described by a nat￾ural language query q. Conventionally, … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages

  1. [5]

    We analyzed na¨ıve adaptations of diffusion models to LASS and identified key lim- itations

    Conclusion We explored the feasibility of training-free LASS by leverag- ing pretrained diffusion models, originally designed for audio generation, for zero-shot source separation. We analyzed na¨ıve adaptations of diffusion models to LASS and identified key lim- itations. To address these challenges, we introduced Diffusion- Guided Mask Optimization, a t...

    work page

  2. [1]

    Computational models aim to replicate this ability through sound separation, isolating target sources from audio mix- tures

    Introduction Humans can focus on specific sounds in complex auditory envi- ronments, a phenomenon known as the cocktail party effect[1]. Computational models aim to replicate this ability through sound separation, isolating target sources from audio mix- tures. Language-queried Audio Source Separation (LASS) has emerged as a flexible solution, allowing us...

    work page

  3. [2]

    A cat is meowing

    Related Works Language-queried Audio Source Separation Early sound separation models achieved success within predefined do- mains [10, 11, 12]. Research has since expanded to universal sound sources using vision [13], audio [14], label [3], and lan- guage queries. The language-based approach is appealing for its accessibility. LASS-Net [2] first introduce...

    arXiv 2025

  4. [3]

    Method 3.1. Language-queried Audio Source Separation Given an audio mixture x composed of multiple source signals {si} and environmental noise e formulated as x = P i si + e, LASS [2] aims to extract a target source s∗ described by a nat- ural language query q. Conventionally, this task is addressed by estimating a mask M (x, q) and applying it to the mix...

    work page

  5. [4]

    Evaluation Benchmarks For evaluation, we use four publicly available text-aligned au- dio datasets and construct artificial mixtures following prior re- search in LASS [3, 4]

    Experiments 4.1. Evaluation Benchmarks For evaluation, we use four publicly available text-aligned au- dio datasets and construct artificial mixtures following prior re- search in LASS [3, 4]. All datasets include both training and test sets. However, as our method is entirely training-free, we exclusively utilize the test set for evaluation. separation m...

    work page

  6. [6]

    Acknowledgements This research was supported by IITP grants (IITP-2025- RS-2020-II201819, IITP-2025-RS-2024-00436857, IITP- 2025-RS-2024-00398115, IITP-2025-RS-2025-02263754, IITP-2025-RS-2025-02304828 ), and the KOCCA grant (RS-2024-00345025 ) funded by the Korea government (MSIT, MOE and MSCT)

    work page 2025

  7. [7]

    The cocktail party problem,

    S. Haykin and Z. Chen, “The cocktail party problem,” Neural computation, vol. 17, no. 9, pp. 1875–1902, 2005

    1902

  8. [8]

    Separate what you describe: Language- queried audio source separation,

    X. Liu, H. Liu, Q. Kong, X. Mei, J. Zhao, Q. Huang, M. D. Plumbley, and W. Wang, “Separate what you describe: Language- queried audio source separation,” in Proceedings of Interspeech 2022, 2022, pp. 1801–1805

    work page 2022

Show all 40 references
  1. [9]

    Clipsep: Learning text-queried sound separation with noisy unlabeled videos,

    H.-W. Dong, N. Takahashi, Y . Mitsufuji, J. McAuley, and T. Berg- Kirkpatrick, “Clipsep: Learning text-queried sound separation with noisy unlabeled videos,” arXiv preprint arXiv:2212.07065 , 2022

    2022 arXiv

  2. [10]

    Separate anything you describe,

    X. Liu, Q. Kong, Y . Zhao, H. Liu, Y . Yuan, Y . Liu, R. Xia, Y . Wang, M. D. Plumbley, and W. Wang, “Separate anything you describe,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2024

    2024

  3. [11]

    Flowsep: Language-queried sound separation with rectified flow matching,

    Y . Yuan, X. Liu, H. Liu, M. D. Plumbley, and W. Wang, “Flowsep: Language-queried sound separation with rectified flow matching,” arXiv preprint arXiv:2409.07614, 2024

    2024 arXiv

  4. [12]

    Soloaudio: Target sound extraction with language-oriented audio diffusion transformer,

    H. Wang, J. Hai, Y .-J. Lu, K. Thakkar, M. Elhilali, and N. De- hak, “Soloaudio: Target sound extraction with language-oriented audio diffusion transformer,” in ICASSP 2025-2025 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  5. [13]

    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,”Proceedings of the International Confer- ence on Machine Learning, pp. 21 450–21 474, 2023

    2023

  6. [14]

    Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,

    J. Xue, Y . Deng, Y . Gao, and Y . Li, “Auffusion: Leveraging the power of diffusion and large language models for text-to-audio generation,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 32, pp. 4700–4712, 2024

    2024

  7. [15]

    Peekaboo: Text to image diffusion models are zero-shot segmentors,

    R. Burgert, K. Ranasinghe, X. Li, and M. S. Ryoo, “Peekaboo: Text to image diffusion models are zero-shot segmentors,”ArXiv, vol. abs/2211.13224, 2022

    2022 arXiv

  8. [16]

    Decoupling magnitude and phase estimation with deep resunet for music source separation,

    Q. Kong, Y . Cao, H. Liu, K. Choi, and Y . Wang, “Decoupling magnitude and phase estimation with deep resunet for music source separation,” in 22nd International Conference on Music Information Retrieval, ISMIR 2021 . International Society for Music Information Retrieval, 2021...

    2021

  9. [17]

    Supervised speech separation based on deep learning: An overview,

    D. Wang and J. Chen, “Supervised speech separation based on deep learning: An overview,” IEEE/ACM transactions on audio, speech, and language processing, vol. 26, no. 10, pp. 1702–1726, 2018

    2018

  10. [18]

    Audio prompt tuning for universal sound separation,

    Y . Liu, X. Liu, Y . Zhao, Y . Wang, R. Xia, P. Tain, and Y . Wang, “Audio prompt tuning for universal sound separation,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1446–1450

    2024

  11. [19]

    The sound of pixels,

    H. Zhao, C. Gan, A. Rouditchenko, C. V ondrick, J. McDermott, and A. Torralba, “The sound of pixels,” in The European Confer- ence on Computer Vision (ECCV), September 2018

    2018

  12. [20]

    Universal source separation with weakly labelled data,

    Q. Kong, K. Chen, H. Liu, X. Du, T. Berg-Kirkpatrick, S. Dubnov, and M. D. Plumbley, “Universal source separation with weakly labelled data,” arXiv preprint arXiv:2305.07447, 2023

    2023 arXiv

  13. [21]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  14. [22]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (I...

    2023

  15. [23]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  16. [24]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” 2021

    2021

  17. [25]

    Scaling rectified flow transformers for high-resolution im- age synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M ¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dock- horn, Z. English, K. Lacey, A. Goodwin, Y . Marek, and R. Rom- bach, “Scaling rectified flow transformers for high-resolution im- age synthesis,” 2024

    2024

  18. [26]

    Audioldm 2: Learn- ing 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: Learn- ing holistic audio generation with self-supervised pretraining,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 32, pp. 2871–2883, 2024

    2024

  19. [27]

    Dreamfusion: Text-to-3d using 2d diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “Dreamfusion: Text-to-3d using 2d diffusion,”arXiv preprint arXiv:2209.14988, 2022

    2022 arXiv

  20. [28]

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

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

    2023

  21. [29]

    Prompt-guided precise audio editing with diffusion models,

    M. Xu, C. Li, D. Su, W. Liang, D. Yu et al. , “Prompt-guided precise audio editing with diffusion models,” arXiv preprint arXiv:2406.04350, 2024

    2024 arXiv

  22. [30]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  23. [31]

    Null-text inversion for editing real images using guided diffusion models,

    R. Mokady, A. Hertz, K. Aberman, Y . Pritch, and D. Cohen-Or, “Null-text inversion for editing real images using guided diffusion models,” arXiv preprint arXiv:2211.09794, 2022

    2022 arXiv

  24. [32]

    Denoising diffusion probabilis- tic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilis- tic models,” Advances in neural information processing systems , vol. 33, pp. 6840–6851, 2020

    2020

  25. [33]

    Vggsound: A large-scale audio-visual dataset,

    H. Chen, W. Xie, A. Vedaldi, and A. Zisserman, “Vggsound: A large-scale audio-visual dataset,” in ICASSP 2020-2020 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2020, pp. 721–725

    2020

  26. [34]

    Audiocaps: Generating captions for audios in the wild,

    C. D. Kim, B. Kim, H. Lee, and G. Kim, “Audiocaps: Generating captions for audios in the wild,” in NAACL-HLT, 2019

    2019

  27. [35]

    Esc: Dataset for environmental sound classifica- tion,

    K. J. Piczak, “Esc: Dataset for environmental sound classifica- tion,” in Proceedings of the 23rd ACM international conference on Multimedia, 2015, pp. 1015–1018

    2015

  28. [36]

    Sdr– half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “Sdr– half-baked or well done?” in ICASSP 2019-2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 626–630

    2019

  29. [37]

    A reference-free metric for language- queried audio source separation using contrastive language-audio pretraining,

    F. Xiao, J. Guan, Q. Zhu, X. Liu, W. Wang, S. Qi, K. Zhang, J. Sun, and W. Wang, “A reference-free metric for language- queried audio source separation using contrastive language-audio pretraining,” CoRR, 2024

    2024

  30. [38]

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

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  31. [39]

    Resunet- a: A deep learning framework for semantic segmentation of re- motely sensed data,

    F. I. Diakogiannis, F. Waldner, P. Caccetta, and C. Wu, “Resunet- a: A deep learning framework for semantic segmentation of re- motely sensed data,” ISPRS Journal of Photogrammetry and Re- mote Sensing, vol. 162, pp. 94–114, 2020

    2020

  32. [40]

    Clotho: an audio cap- tioning dataset,

    K. Drossos, S. Lipping, and T. Virtanen, “Clotho: an audio cap- tioning dataset,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 736– 740

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.