Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Mel-Refine: A Plug-and-Play Approach to Refine Mel-Spectrogram in Audio Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read By reweighting high- and low-frequency components of the U-Net feature maps during inference, Mel-Refine improves text-to-audio quality without retraining, cutting Tango2's FD from 37.48 to 28.13 and raising listener preference from 36 to…

desk verdict A plausible no-training U-Net frequency hack for TTA, but the 25% headline is a fitted number because the knobs were tuned on the same evaluation subset used to report gains. read the letter →

arxiv 2412.08577 v1 pith:KT3GMEDB submitted 2024-12-11 cs.SD cs.MMeess.AS

Hongming Guo , Ruibo Fu , Yizhong Geng , Shuai Liu , Shuchen Shi , Tao Wang , Chunyu Qiang , Chenxing Li
show 4 more authors
Ya Li Zhengqi Wen Yukun Liu Xuefei Liu
This is my paper · ORCID
classification cs.SDcs.MMeess.AS
keywords text-to-audiogenerationMel-spectrogramdiffusionmodelU-Netfrequency-domainrefinementplug-and-playinferenceTango2audioqualitymetrics
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

Text-to-audio models that generate Mel-spectrograms with a U-Net-based diffusion backbone tend to blur or muffle complex sound scenes because they fail to render fine texture. Mel-Refine is a training-free, inference-only intervention that rebalances how much the U-Net's skip connections and backbone contribute at high and low frequencies: it amplifies high-frequency components in skip connections, applies a structure-aware amplification to the backbone, and suppresses high-frequency components in the backbone. On Tango, Tango2, and MusTango, these simple weight changes improve objective audio-quality metrics and subjective overall-quality scores. The point is that a model's quality ceiling is not fixed after training; the frequency structure of its own feature maps during inference contains recoverable detail.

What carries the argument

The central object is the U-Net decoder feature map split by Fourier transform into high- and low-frequency regions. A scaling matrix $\beta$ leaves the central low-frequency band unchanged while scaling the rest by $s$, applied to skip-connection features; the backbone gets a structure-aware scaling based on channel-wise means (controlled by $m$) followed by a second Fourier-domain mask that attenuates high frequencies with $b$. These operations are applied only in the first two decoder blocks, where features are densest. This machinery turns the empirical finding into a concrete inference-time recipe: reweight, rather than retrain, to recover high-frequency texture.

What would settle it

Run Mel-Refine with the paper's published parameter settings on a held-out set of AudioCaps captions never used in the grid search, keeping all other inference settings fixed; if FD, FAD, and KL show little or no improvement over baseline, or listener OVL does not rise, the reported gains are tuning artifacts rather than a general inference-time effect.

Watch

Extended reading notes

Core claim

This paper establishes that the frequency content of U-Net feature maps has separable roles in Mel-spectrogram generation: the high-frequency parts of both skip connections and backbone determine texture and edge detail, while the low-frequency backbone content drives the denoising process. Based on Fourier analysis of intermediate features, Mel-Refine adjusts these components during inference, multiplying skip-connection high frequencies by $s>1$, scaling backbone features with a structure-aware map controlled by $m>1$, and attenuating backbone high frequencies by $b<1$, all within the first two decoder blocks. The authors show that this restores detail and coherence in generated audio, with the largest gains when the base model has lost generation quality through preference alignment, and report consistent improvements in FD, FAD, KL, and listener OVL across three models. The intended contribution is a plug-and-play refinement stage that requires no additional training or fine-tuning and is compatible with any diffusion-based TTA architecture.

Load-bearing premise

The gains rest on the assumption that the five per-model parameters chosen by grid search on the evaluation subset generalize to audio beyond that subset, rather than merely fitting the test data.

Editorial extensions

If this is right

  • Any diffusion-based text-to-audio model with a U-Net can adopt Mel-Refine at inference time, with no retraining or fine-tuning, and improve FD, FAD, and subjective overall quality.
  • The frequency-decomposition result implies that amplifying low-frequency components in the U-Net backbone, rather than high-frequency ones, is what causes over-smoothing or degraded denoising in Mel-spectrogram generation.
  • The largest gains appear on models whose alignment fine-tuning traded away generation quality, so Mel-Refine offers a way to recover quality after alignment without retraining.
  • Because only the first two decoder blocks are modified, the added inference cost is small enough to make the method practical as a default post-processing step.

Reading between the lines

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

  • The paper does not test whether the same frequency-reweighting recipe transfers to other U-Net-based latent diffusion domains, such as image or video generation, but the mechanism is stated in terms of generic feature maps rather than audio-specific structure.
  • A stricter evaluation would fix the five parameters on a training subset and measure on a separate held-out caption split, since the paper's grid search and evaluation share the same subset.
  • The method could plausibly be combined with other inference-time interventions or applied at more than the first two decoder blocks; the paper does not explore whether later decoder stages would help or hurt.
  • If the proposed explanation is correct, the benefit should scale with prompt complexity: prompts with many sound events should show larger gains than simple single-event prompts, which is a directly testable prediction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 Mel-Refine, a training-free inference-time method to improve the quality of Mel-spectrograms generated by latent-diffusion text-to-audio (TTA) models. The method applies frequency-domain scaling to the high-frequency components of U-Net skip connections and backbone features in the first two decoder blocks, controlled by five scalar parameters (s1, s2, b1, b2, m). The authors report experiments on Tango, Tango2, and MusTango, showing improvements in FD, FAD, KL, and subjective OVL. The central claim is that high-frequency components in skip connections enhance texture, while low-frequency backbone components are critical for denoising, and that adjusting these weights during inference improves audio fidelity.

Significance. The idea of reweighting U-Net frequency components at inference is timely and draws on a known phenomenon from image diffusion models (FreeU). If the reported improvements were robust, the method would be a useful, training-free plug-in for LDM-based TTA models. The paper's strengths are the concrete mechanism and the qualitative spectrogram comparisons; the weaknesses are the in-sample parameter selection and lack of statistical rigor, which affect the credibility of all quantitative claims.

major comments (4)
  1. [III-A, III-C] The parameter-selection and evaluation protocol is circular. In Section III-C, the five parameters (s1, s2, b1, b2, m) are determined by a two-stage grid search, and the resulting settings are used for the results in Table II. Section III-A states that evaluation is performed on 'only a subset of the data' rather than the full test set. Because the same subset is used both for tuning and for reporting, the improvements, including the headline Tango2 FD reduction from 37.48 to 28.13, are in-sample fits rather than predictive results. The authors must use a separate validation set for parameter search and a disjoint test set for final evaluation, or perform nested cross-validation, and they should report the subset size and the number of parameter configurations tried.
  2. [III-D, Table II] The paper reports no uncertainty quantification. All objective metrics are single numbers, so the small differences for Tango (FD 28.36 vs 27.78) and MusTango (FAD 1.67 vs 1.61) cannot be distinguished from noise. Similarly, the subjective OVL evaluations are based on only 10 evaluators and 30 clips, with no description of whether listening tests were blinded, randomized, or analyzed for significance. Confidence intervals, standard errors, or significance tests are needed before the claimed improvements can be accepted.
  3. [II-B, Eq. (4)] The frequency mask in Eq. (4) defines the 'central region' as low-frequency, with x in [1/4W, 3/4W] and y in [1/4H, 3/4H]. In a standard 2D FFT without fftshift, the low-frequency (DC) components are located at the corners, not the center. If the authors are using a library that does not shift the spectrum, this mask would amplify the wrong region. The manuscript must specify the exact FFT implementation and whether fftshift/ifftshift is used; otherwise, the mechanism described (enhancing high frequencies in skip connections, attenuating them in the backbone) may be operating on the wrong frequency bands.
  4. [III-E, Table III] The ablation study does not provide independent evidence for the contribution of each module. It removes one module at a time while keeping the parameters tuned for the full method, on the same evaluation subset used for tuning. A proper ablation should re-tune or at least evaluate on a held-out set, and it should report whether parameter re-optimization changes the pattern. The current table only shows that the method's components are interdependent under a particular tuning, not that each component is necessary.
minor comments (5)
  1. [III-A] The size of the evaluation subset is not stated; knowing how many clips are used is essential for interpreting the stability of FD/FAD/KL estimates.
  2. [II-A, Table I] The FAD value of 9.58 for 'Backbone Low-frequency Attenuate' is far outside the range of the other entries; please verify this number and add a sentence of explanation.
  3. [II-B, Eq. (6)] The min and max in the normalization of α are not defined; it should be stated whether they are taken over the spatial dimensions of each channel or over the channel dimension.
  4. [I] The claim of full compatibility with 'any diffusion-based TTA architecture' is not supported by the experiments, which cover only LDM-based models; a more cautious statement would be appropriate.
  5. [Abstract] The '25% boost' is not defined; if it refers to the relative decrease in FD, this should be stated explicitly, along with the fact that it is a single-metric, single-model result.

Circularity Check

1 steps flagged · score 7.0 of 10

The five per-model parameters are tuned by grid search on the same evaluation subset whose metrics are then reported, making the headline Tango2 improvement an in-sample fitted result rather than a held-out prediction.

  1. fitted input called prediction [Section III-C (Model Setting Searching); Table II; Section III-A (subset evaluation)]
    "Afterward, a grid search was conducted to determine the optimal parameters. The parameters provided in the subsequent results are the optimal ones obtained through experiments."

    In Table II the reported Mel-Refine FD/FAD/KL values (e.g., Tango2 FD 37.48 to 28.13) are computed after a grid search chosen specifically to optimize these metrics. Section III-C describes the search: 'Afterward, a grid search was conducted to determine the optimal parameters. The parameters provided in the subsequent results are the optimal ones obtained through experiments.' Section III-A says the evaluation uses 'only a subset of the data, rather than testing the entire dataset.' Therefore the same subset serves both as the tuning set and the report set. With five free parameters per model, the search can exploit noise in the small subset, so the improvement is a fitted value, not an independent prediction of Mel-Refine's effect.

full rationale

The only substantive circularity is in the evaluation protocol, not in the method's derivation. Sections II-A and II-B derive Mel-Refine from controlled component experiments (Table I) and from FreeU; those experiments are evidence for the mechanism, though the coefficient choices are not specified. However, the headline claim ('boosts performance metrics of Tango2 by 25%', abstract) rests on Table II, which is produced with parameters selected in Section III-C. Section III-C states: 'Afterward, a grid search was conducted to determine the optimal parameters. The parameters provided in the subsequent results are the optimal ones obtained through experiments.' The evaluation in Section III-A uses 'only a subset of the data.' Thus the reported FD/FAD/KL values are the outcome of optimizing five per-model parameters (s1, s2, b1, b2, m) on the same subset on which they are then reported. The improvement is an in-sample selection result, not an independent estimate of Mel-Refine's effect; the same metrics on the same subset after tuning are statistically forced to look better than the untuned baseline. The ablation (Table III) uses the already-tuned configuration, so it does not provide held-out confirmation. No confidence intervals or significance tests are reported. This is a fitted-input-called-prediction pattern. It does not make the paper definitionally circular, since the mechanism and the U-Net frequency analysis have independent content, and a properly held-out evaluation could in principle confirm the method. Hence the score is 7 rather than higher.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several domain assumptions and hand-chosen design choices. The method's numerical results are not derived from first principles; they depend on free parameters tuned to the evaluation set. No new physical or conceptual entities are introduced.

free parameters (5)
  • skip-connection high-frequency scaling factors s1, s2 = e.g., Tango2: s1=1.4, s2=1.2
    Chosen by grid search on the evaluation subset to optimize FD, FAD, and KL; values vary per model.
  • backbone high-frequency attenuation factors b1, b2 = e.g., Tango2: b1=0.5, b2=0.1
    Chosen by grid search on the evaluation subset to optimize FD, FAD, and KL.
  • backbone structure scaling factor m = e.g., Tango2: m=2.5
    Chosen by grid search on the evaluation subset to optimize FD, FAD, and KL.
  • central region bounds for low frequency = x in [1/4W, 3/4W], y in [1/4H, 3/4H]
    Hand-chosen in Equation 4; no sensitivity analysis is provided.
  • number of decoder blocks adjusted (first two) = 2 blocks
    Hand-chosen based on the claim that features become sparser in later decoder stages; no experiment supports this specific choice.
assumptions (4)
  • domain assumption Frequency decomposition via FFT and separation into low/high components is meaningful for Mel-spectrogram U-Net features.
    Used throughout Section II; no evidence is given that Fourier-domain manipulation preserves the semantic structure of the features.
  • ad hoc to paper FreeU's structure-aware backbone scaling transfers to audio U-Nets.
    Adopted from FreeU without re-derivation or validation in the Mel-spectrogram domain beyond the reported experiments.
  • domain assumption Evaluation metrics FD, FAD, and KL are valid proxies for audio generation quality.
    Used as objective metrics in Section III-B; standard in prior TTA work but their limitations are not discussed.
  • domain assumption The selected random subset and fixed seed are representative of the test distribution.
    Evaluation in Section III-A uses a subset of AudioCaps and MusicBench with one caption per clip; no confidence intervals or multiple-seed analysis are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mel-Refine: A Plug-and-Play Approach to Refine Mel-Spectrogram in Audio Generation." pith.science (2026). https://pith.science/paper/KT3GMEDB

@misc{pith2026241208577,
  author       = {Pith},
  title        = {Pith review of: Mel-Refine: A Plug-and-Play Approach to Refine Mel-Spectrogram in Audio Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KT3GMEDB}},
  note         = {Machine review of arXiv:2412.08577}
}
read the original abstract

Text-to-audio (TTA) model is capable of generating diverse audio from textual prompts. However, most mainstream TTA models, which predominantly rely on Mel-spectrograms, still face challenges in producing audio with rich content. The intricate details and texture required in Mel-spectrograms for such audio often surpass the models' capacity, leading to outputs that are blurred or lack coherence. In this paper, we begin by investigating the critical role of U-Net in Mel-spectrogram generation. Our analysis shows that in U-Net structure, high-frequency components in skip-connections and the backbone influence texture and detail, while low-frequency components in the backbone are critical for the diffusion denoising process. We further propose ``Mel-Refine'', a plug-and-play approach that enhances Mel-spectrogram texture and detail by adjusting different component weights during inference. Our method requires no additional training or fine-tuning and is fully compatible with any diffusion-based TTA architecture. Experimental results show that our approach boosts performance metrics of the latest TTA model Tango2 by 25\%, demonstrating its effectiveness.

Figures

Figures reproduced from arXiv: 2412.08577 by the authors.

Figure 1
Figure 1. Comparison of Mel-spectrograms between simple and complex audio [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. This figure illustrates the impact of altering weights at different [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The whole framework of Mel-Refine Here, ⊙ denotes element-wise multiplication, and β(x, y) is a scaling matrix defined as: β(x, y) = ( 1, (x, y) ∈ central region s, otherwise (4) Term central region corresponds to the low-frequency components in the frequency domain: x ∈ [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The left Mel-spectrogram shows the original output from the model, [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 12 canonical work pages

  1. [1]

    Hmm- based text-to-audio-visual speech synthesis

    S. Sako, K. Tokuda, T. Masuko, T. Kobayashi, and T. Kitamura, “Hmm- based text-to-audio-visual speech synthesis.” in INTERSPEECH, vol. 25, 2000, p. 28

    work page 2000

  2. [2]

    Echoaudio: Efficient and high-quality text-to-audio generation with minimal inference steps,

    H. Liu, R. Huang, Y . Liu, H. Cao, J. Wang, X. Cheng, S. Zheng, and Z. Zhao, “Echoaudio: Efficient and high-quality text-to-audio generation with minimal inference steps,” in ACM Multimedia 2024 , 2024

    work page 2024

  3. [3]

    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,” arXiv preprint arXiv:2401.01044, 2024

    arXiv 2024

  4. [4]

    Make-an-audio: Text-to-audio generation with prompt- enhanced diffusion models,

    R. Huang, J. Huang, D. Yang, Y . Ren, L. Liu, M. Li, Z. Ye, J. Liu, X. Yin, and Z. Zhao, “Make-an-audio: Text-to-audio generation with prompt- enhanced diffusion models,” in International Conference on Machine Learning. PMLR, 2023, pp. 13 916–13 932

    2023

  5. [5]

    Make-an-audio 2: Temporal-enhanced text-to- audio generation,

    J. Huang, Y . Ren, R. Huang, D. Yang, Z. Ye, C. Zhang, J. Liu, X. Yin, Z. Ma, and Z. Zhao, “Make-an-audio 2: Temporal-enhanced text-to- audio generation,” arXiv preprint arXiv:2305.18474 , 2023

    arXiv 2023

  6. [6]

    DiffAVA: Personalized Text-to-Audio Generation with Visual Alignment

    S. Mo, J. Shi, and Y . Tian, “Diffava: Personalized text-to-audio genera- tion with visual alignment,” arXiv preprint arXiv:2305.12903 , 2023

    work page Pith review arXiv 2023

  7. [7]

    Audiolm: a language modeling approach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Shar- ifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasacchi et al., “Audiolm: a language modeling approach to audio generation,” IEEE/ACM transac- tions on audio, speech, and language processing, vol. 31, pp. 2523–2533, 2023

    work page 2023

  8. [8]

    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 Transactions on Audio, Speech, and Language Processing , 2024

    2024

Show all 29 references
  1. [9]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyen, A. Bapna, Z. Bor- sos, F. d. C. Quitry, P. Chen, D. E. Badawy, W. Han, E. Kharitonov et al., “Audiopalm: A large language model that can speak and listen,” arXiv preprint arXiv:2306.12925 , 2023

    2023 arXiv

  2. [10]

    Pengi: An audio language model for audio tasks,

    S. Deshmukh, B. Elizalde, R. Singh, and H. Wang, “Pengi: An audio language model for audio tasks,” Advances in Neural Information Processing Systems, vol. 36, pp. 18 090–18 108, 2023

    2023

  3. [11]

    Uniaudio 1.5: Large language model-driven audio codec is a few-shot audio task learner,

    D. Yang, H. Guo, Y . Wang, R. Huang, X. Li, X. Tan, X. Wu, and H. Meng, “Uniaudio 1.5: Large language model-driven audio codec is a few-shot audio task learner,” arXiv preprint arXiv:2406.10056 , 2024

    2024 arXiv

  4. [12]

    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,” in The Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/ fo...

    2023

  5. [13]

    Simple and controllable music generation,

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y . Adi, and A. D ´efossez, “Simple and controllable music generation,” in Thirty-seventh Conference on Neural Information Processing Systems ,

  6. [14]

    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 Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 1720–1733, 2023

    2023

  7. [15]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  8. [16]

    Lafma: A latent flow matching model for text- to-audio generation,

    W. Guan, K. Wang, W. Zhou, Y . Wang, F. Deng, H. Wang, L. Li, Q. Hong, and Y . Qin, “Lafma: A latent flow matching model for text- to-audio generation,” arXiv preprint arXiv:2406.08203 , 2024

    2024 arXiv

  9. [17]

    Text-to-audio gen- eration using instruction guided latent diffusion model,

    D. Ghosal, N. Majumder, A. Mehrish, and S. Poria, “Text-to-audio gen- eration using instruction guided latent diffusion model,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 3590–3598

    2023

  10. [18]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2308.05734

    2024 arXiv

  11. [19]

    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,” in Proceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. ...

    2023

  12. [20]

    Mustango: Toward controllable text-to-music generation,

    J. Melechovsky, Z. Guo, D. Ghosal, N. Majumder, D. Herremans, and S. Poria, “Mustango: Toward controllable text-to-music generation,”

  13. [21]

    Tango 2: Aligning diffusion-based text-to-audio generative models through direct preference optimization,

    N. Majumder, C.-Y . Hung, D. Ghosal, W.-N. Hsu, R. Mihalcea, and S. Poria, “Tango 2: Aligning diffusion-based text-to-audio generative models through direct preference optimization,” in ACM Multimedia 2024 , 2024. [Online]. Available: https://openreview.net/ forum?id=7lqptq5dLG

    2024

  14. [22]

    Freeu: Free lunch in diffusion u-net,

    C. Si, Z. Huang, Y . Jiang, and Z. Liu, “Freeu: Free lunch in diffusion u-net,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 4733–4743

    2024

  15. [23]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 ...

    2015

  16. [24]

    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 Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), ...

    2019

  17. [25]

    Audio set: An ontology and human- labeled dataset for audio events,

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human- labeled dataset for audio events,” in 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2017,...

    2017

  18. [26]

    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,” arXiv preprint arXiv:1812.08466, 2018

    2018 arXiv

  19. [27]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  20. [2023]

    Available: https://openreview.net/forum?id=jtiQ26sCJi

    [Online]. Available: https://openreview.net/forum?id=jtiQ26sCJi

  21. [2024]

    Available: https://arxiv.org/abs/2311.08355

    [Online]. Available: https://arxiv.org/abs/2311.08355

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.