Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Few-step Adversarial Schr\"{o}dinger Bridge for Generative Speech Enhancement

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 This paper shows that adding GAN-style adversarial training to a Schrödinger Bridge for speech enhancement reduces inference to a single step while outperforming 50-step baselines on full-band denoising and dereverberation.

desk verdict Useful empirical paper: one-step SB-GAN hybrid for full-band speech enhancement works, but the abstract oversells the dereverberation result. read the letter →

arxiv 2506.01460 v1 pith:6PVWC5CL submitted 2025-06-02 cs.SD eess.AS

Seungu Han , Sungho Lee , Juheon Lee , Kyogu Lee This is my paper
classification cs.SDeess.AS
keywords speechenhancementSchrödingerbridgegenerativeadversarialnetworksdiffusion-GANhybridone-stepinferencedenoisingdereverberationfull-band
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

The paper tries to establish that generative speech enhancement need not be slow: by fusing a Schrödinger Bridge with GAN-based distribution matching, high-quality full-band denoising and dereverberation can be done in one inference step. This matters because existing diffusion and Schrödinger Bridge models typically require more than 50 iterative steps, which blocks real-time use. The authors claim their single-step model outperforms 50-step baselines on the EARS-WHAM and EARS-Reverb benchmarks, with the largest gains in low-SNR conditions. The result, if correct, makes generative enhancement competitive with fast predictive models while retaining the generative model's ability to repair missing information.

What carries the argument

The central object is a conditional Schrödinger Bridge with closed-form Gaussian marginals, discretized into N=4 steps, combined with a UFOGen-style adversarial parameterization. The generator produces a clean estimate x'_0 = G_theta(x_tn, y, t_n), then a sample x'_{t_{n-1}} is drawn from the forward Gaussian q(x_{t_{n-1}}|x'_0, y); a time-conditioned discriminator distinguishes this sample from the true intermediate q(x_{t_{n-1}}|x_0, y), while a reconstruction loss ties x'_0 to the true clean speech.

What would settle it

Take the trained single-step SB-UFOGen and the 50-step SB baseline and run both on EARS-WHAM utterances at -5 dB input SNR; if the single-step model does not match or beat the baseline on PESQ and SI-SDR, the central claim is false.

Watch

Extended reading notes

Core claim

The paper discovers that the Schrödinger Bridge's robustness to fewer sampling steps breaks down on full-band speech under low-SNR conditions, and that this degradation can be repaired by replacing the coarse-step denoising distribution with an adversarially trained generator. The resulting model, SB-UFOGen, matches the intermediate denoising distribution at four discrete steps and, with a single step, outperforms the 50-step Schrödinger Bridge baseline on both denoising and dereverberation. The authors further show that single-step inference acts as a direct predictive mapping, yielding higher intrusive-metric scores than four-step inference in some settings.

Load-bearing premise

The result depends on the assumption that the UFOGen adversarial-training recipe, developed for image diffusion, still approximates the true denoising distribution when grafted onto a four-step Schrödinger Bridge for speech; if that match fails on unseen low-SNR or reverberant conditions, the one-step quality claim would break.

Editorial extensions

If this is right

  • Single-step inference runs at 0.046 seconds of processing per second of audio on an RTX A6000, roughly 45 times faster than the 50-step SB baseline, making generative enhancement practical for real-time applications.
  • The adversarial correction preserves quality when the number of steps is reduced, especially in low-SNR conditions where the baseline SB model loses over 9 dB SI-SDR from 32 steps to 1 step.
  • The same training recipe transfers from denoising to dereverberation with only the reconstruction loss adjusted (adding a multi-scale mel loss), and single-step SB-UFOGen matches or beats the 50-step SB baseline on both tasks.
  • Even when the PESQ metric is optimized, single-step SB-UFOGen outperforms the 50-step SB baseline, although optimizing PESQ trades off non-intrusive quality metrics.
  • The framework is not tied to the VE Schrödinger Bridge: any diffusion process that can be defined as a Markov chain, such as the OUVE process, can use the same adversarial training scheme.

Reading between the lines

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

  • If one step suffices, the trained model behaves like a deterministic regressor at inference; the generative aspect of the system may matter more for how the distribution is learned during training than for sampling diversity at test time.
  • The low-SNR collapse of the baseline SB suggests that the Gaussian approximation of the denoising distribution fails precisely when the target distribution is most multimodal, and GAN matching is one remedy; other distribution-matching losses might yield similar robustness.
  • Because the formulation only requires Gaussian transition kernels, the recipe should transfer to other Schrödinger Bridge variants and to image or audio restoration tasks such as super-resolution, which the paper names as future work.
  • A practical test is whether the single-step margin persists under more severe degradations than those in EARS-WHAM and EARS-Reverb, such as -5 dB SNR or highly realistic room reverberation.
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 SB-UFOGen, a generative speech enhancement method that integrates a Schrödinger Bridge (SB) formulation with the UFOGen adversarial training objective, so that the reverse process can be run with very few steps (down to one). The model is trained on full-band EARS-WHAM and EARS-Reverb datasets, and the authors report results for denoising and dereverberation in terms of PESQ, SI-SDR, ESTOI, DNSMOS, SIGMOS, and RTF. The central claim is that SB-UFOGen outperforms existing baselines, including 50-step SB models, even with a single inference step, and that it remains robust at low SNR where standard SB models degrade sharply.

Significance. If the empirical results hold, the paper addresses a real bottleneck of Schrödinger Bridge speech enhancement: the need for many reverse-time sampling steps. The low-SNR analysis in Figure 1 is a useful addition, and the reported RTF improvement (single-step SB-UFOGen at 0.046 proc/s versus 2.112 proc/s for 50-step SB) is practically relevant. The use of full-band datasets and both intrusive and non-intrusive metrics is a strength. However, the headline claim is currently stronger than the evidence: the paper's own Table 3 shows a PESQ deficit for single-step SB-UFOGen relative to 50-step SB on dereverberation, and no significance tests or error bars are reported anywhere. These issues must be addressed before the central claim can be accepted as stated.

major comments (4)
  1. [Abstract and §4.3 (Table 3)] The abstract states that the model 'outperforms existing baselines, even with a single inference step, in both denoising and dereverberation tasks,' but Table 3 reports single-step SB-UFOGen PESQ 3.36 versus the 50-step SB baseline PESQ 3.41 on EARS-Reverb. Since PESQ is the most widely used intrusive quality metric, this is a direct counterexample to the universal outperformance claim. The claim should be qualified (e.g., 'outperforms on most metrics, with a large SI-SDR gain, while remaining comparable on PESQ') or supported by paired significance tests showing that the 0.05 PESQ gap is negligible.
  2. [Tables 1–3] No error bars, confidence intervals, or significance tests are reported. Many of the reported favorable comparisons are extremely small (for example, Table 1 shows DNSMOS 3.88 vs. 3.87 and ESTOI 0.74 vs. 0.73 relative to the 50-step SB baseline; Table 3 shows SIGMOS 3.33 vs. 3.37). Without variance estimates or paired tests, the reader cannot determine whether these differences are real or within training/evaluation noise. Please report standard deviations over multiple seeds or at least paired significance tests on the same test utterances.
  3. [§4.2–§4.3, Tables 1–3] The SB baselines are not drawn from a single consistent source: Table 1 uses SB from [17], while Tables 2 and 3 use SB from [4]. The PESQ-loss variant in Table 2 yields a much stronger SB baseline (PESQ 3.08) than the non-PESQ-loss SB in Table 1 (PESQ 2.32), yet the paper does not clarify which numbers come from the authors' own runs under identical training and evaluation conditions and which are taken from the cited papers. This makes the 'outperforms existing baselines' claim difficult to verify, since baseline strength varies across tables.
  4. [§3.2 and Figures 1–2] The model is trained with N=4 discretization steps, but Figures 1 and 2 show results for SB-UFOGen at 1, 2, 4, 8, 16, and 32 steps. The paper does not explain how a model trained with N=4 is sampled at other step counts, or whether the same generator is reused with different discretizations. Without this explanation, the comparison of step-sensitivity between SB-UFOGen and the SGMSE+/SB baselines is not apples-to-apples and should be clarified.
minor comments (5)
  1. [Table 3] The metric columns are labeled 'SIGMOS' and 'MOS Reverb', but the text says 'SIGMOS Reverb is reported instead of the DNSMOS'. Please define both metrics clearly and state which one is the reverb-specific MOS.
  2. [§3.2] The sentence 'so our discriminator operates on white Gaussian noise' is confusing: the discriminator input is the intermediate sample x_{t_{n-1}} or its generated estimate, not noise. Please rephrase to describe the actual input representation.
  3. [§2.2, Eqs. (7) and (8)] The notation for the adversarial divergence is inconsistent between Eq. (7), which writes D_adv(q(x_{t_{n-1}}|x_{t_n}) || p_theta(...)), and Eq. (8), which writes D_adv(q(x_{t_{n-1}}) | p_theta(...)). Please make the notation uniform.
  4. [Figure 1] The low-SNR analysis would be easier to interpret if the figure reported the number of test utterances in each SNR bin, and ideally error bars, because the claimed 'drastic' degradation is based on aggregate statistics of unknown sample size.
  5. [§4.2, RTF measurement] The real-time factor is reported as 2.112 proc/s and 0.046 proc/s; please state explicitly that higher values mean faster processing and specify the hardware and batch size used for the timing measurement.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: the training loss is a direct adversarial SB objective and evaluations are on held-out test sets; the abstract overclaim vs. Table 3 is a correctness/statistics concern, not circularity.

full rationale

The paper's derivation chain is self-contained in the sense relevant to circularity: the Schrödinger Bridge equations (Eqs. 1–6) are taken from prior work, and the proposed method's losses in Algorithm 1 and Section 3.1 are a standard combination of adversarial distribution matching, conditional Gaussian sampling, and an l2/l1 reconstruction term. No parameter is fitted to the test set and then renamed a prediction; the reported PESQ, SI-SDR, ESTOI, DNSMOS, and SIGMOS values are computed on the held-out EARS-WHAM and EARS-Reverb test sets. The only self-citation is the unrelated prior work on neural audio upsampling (Ref. 37), and it is not load-bearing. The skeptical concern that the abstract's 'outperforms existing baselines ... in both denoising and dereverberation tasks' is contradicted by Table 3's PESQ result (SB-UFOGen 1-step 3.36 vs. SB 50-step 3.41) is a legitimate empirical-claim issue, but it is not a circularity: the metric values are not constructed from the claim, and the single-step method still improves on SI-SDR. No equation reduces to its own input, no fitted value is presented as a prediction, and no uniqueness or ansatz is imported solely from the authors' prior work. Thus the circularity score is 1, reflecting only a minor non-load-bearing self-citation and no meaningful circular step.

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

The model builds on closed-form Gaussian SB from [4], UFOGen's adversarial objective from [20], and a VE noise schedule from [4]. It introduces no new physical entities, but inherits several assumptions without re-deriving them. The central empirical claim does not depend on tuning exotic parameters beyond standard loss weights.

free parameters (5)
  • N (discretization steps) = 4
    Number of reverse-process steps set to 4, following [18-20]; not tuned here but affects computational cost versus quality.
  • lambda_recon = 100
    Reconstruction loss weight in the generator objective (Section 3.2).
  • alpha (L1 loss weight) = 1e-3
    Weight for time-domain L1 term in the reconstruction loss (Section 3.2).
  • mel loss weight (dereverberation) = 0.01
    Additional multi-scale mel loss weight for dereverberation (Section 3.2).
  • VE schedule parameters (c, k) = from [4]
    Noise schedule hyperparameters inherited from prior SB work [4]; not independently derived in this paper.
assumptions (4)
  • domain assumption Tractable Gaussian solution to SB with linear drift and Gaussian boundary conditions (Eq. 4-6 from [4])
    The entire training relies on closed-form SB marginals q(x_t|x0,y) from prior work [4]; if these are inaccurate for full-band speech, the method fails.
  • domain assumption Markov property of the SB process allows discrete-time GAN matching
    The paper extends UFOGen by assuming the discretized SB reverse process is a Markov chain and that adversarial matching of q(x_{t_{n-1}}|x'_0,y) is a valid objective.
  • standard math VE noise schedule and hyperparameters from [4] are appropriate for full-band speech
    The model adopts the noise schedule and parameterization from [4] without re-derivation for the EARS datasets.
  • ad hoc to paper UFOGen's adversarial training objective transfers from image diffusion to speech SB
    The central method assumes the success of Diffusion-GAN hybrids on images transfers to the speech enhancement SB setting; no proof is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-step Adversarial Schr\"{o}dinger Bridge for Generative Speech Enhancement." pith.science (2026). https://pith.science/paper/6PVWC5CL

@misc{pith2026250601460,
  author       = {Pith},
  title        = {Pith review of: Few-step Adversarial Schr\"odinger Bridge for Generative Speech Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PVWC5CL}},
  note         = {Machine review of arXiv:2506.01460}
}
read the original abstract

Deep generative models have recently been employed for speech enhancement to generate perceptually valid clean speech on large-scale datasets. Several diffusion models have been proposed, and more recently, a tractable Schr\"odinger Bridge has been introduced to transport between the clean and noisy speech distributions. However, these models often suffer from an iterative reverse process and require a large number of sampling steps -- more than 50. Our investigation reveals that the performance of baseline models significantly degrades when the number of sampling steps is reduced, particularly under low-SNR conditions. We propose integrating Schr\"odinger Bridge with GANs to effectively mitigate this issue, achieving high-quality outputs on full-band datasets while substantially reducing the required sampling steps. Experimental results demonstrate that our proposed model outperforms existing baselines, even with a single inference step, in both denoising and dereverberation tasks.

Figures

Figures reproduced from arXiv: 2506.01460 by the authors.

Figure 2
Figure 2. EARS-Reverb results across different sampling steps. incorporating the PESQ loss leads to a decline in other metrics, particularly non-intrusive metrics such as DNSMOS and SIG￾MOS. This is expected, as PESQ loss only prioritizes PESQ, increasing the risk of overfitting to this specific metric [36]. 4.3. Dereverberation Results We also trained and tested the models using the full-band EARS-Reverb dataset [21] for the… view at source ↗
Figure 1
Figure 1. Results on EARS-WHAM test set across different sam￾pling steps and SNR ranges [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [17]

    Diffusion schr¨odinger bridge with applications to score-based generative modeling,

    V . De Bortoli, J. Thornton, J. Heng, and A. Doucet, “Diffusion schr¨odinger bridge with applications to score-based generative modeling,” inNeurIPS, vol. 34, 2021, pp. 17 695–17 709

    work page 2021

  2. [4]

    Experimental Setup We trained our model with a batch size of 8, and the optimizer was AdamW [26] with a learning rate of10 −4

    Evaluation 4.1. Experimental Setup We trained our model with a batch size of 8, and the optimizer was AdamW [26] with a learning rate of10 −4. We use the exponential moving average (EMA) of the weights with a de- cay of 0.999 [27]. The average SI-SDR of 20 randomly se- lected validation examples is logged during training, and the best-performing model is ...

    work page

  3. [1]

    Introduction Deep generative models have been employed for various speech enhancement and restoration tasks [1–6]. This approach is motivated by the inherent nature of the problem: under severe conditions, such as low signal-to-noise ratio (SNR) scenarios, some information is lost, making it necessary to generate the missing parts [7]. Furthermore, unlike...

    work page

  4. [2]

    Background 2.1. Schr ¨odinger Bridge for Speech Enhancement The Schr ¨odinger Bridge (SB) problem [10] considers the opti- mization of path measures with boundary conditions: min Q∈P[p0 ,pT ] DKL(Q||P).(1) Here,Pdenotes a reference measure, which is generally set to the path measure of the forward stochastic differential equation (SDE):dx t =f(t,x t)dt+g(...

    work page Pith review arXiv 2025

  5. [3]

    Adversarial Schr ¨odinger Bridge We combine the Schr¨odinger Bridge (SB) for speech enhance- ment with GANs

    Methods 3.1. Adversarial Schr ¨odinger Bridge We combine the Schr¨odinger Bridge (SB) for speech enhance- ment with GANs. Since the SB is defined by a pair of SDE, it inherently satisfies the Markov property. Thus, the discrete ver- sion is obtained by discretizing the continuous process and then adapting it into the Diffusion-GAN Hybrids, in the same way...

    work page

  6. [5]

    However, both approaches are still constrained by the iterative reverse process

    Conclusion Diffusion models have gained attention in speech enhancement for their generative capabilities, while SB models have recently emerged as a promising alternative. However, both approaches are still constrained by the iterative reverse process. Our in- vestigation reveals that SB models suffer notable performance decline, particularly under low-S...

    work page

  7. [6]

    RS-2024-00461617, 50%), and the Institute of Information & Communications Technology Planning & Eval- uation (IITP) grant funded by the Korea government (MSIT) [No

    Acknowledgements This work was supported by the National Research Founda- tion of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00461617, 50%), and the Institute of Information & Communications Technology Planning & Eval- uation (IITP) grant funded by the Korea government (MSIT) [No. RS-2022-II220641, 40%], [No. RS-2021-II211343, Ar...

    work page 2024

  8. [7]

    Conditional diffusion probabilistic model for speech en- hancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech en- hancement,” inIEEE ICASSP, 2022, pp. 7402–7406

    work page 2022

Show all 43 references
  1. [8]

    Universal speech enhancement with score-based diffusion,

    J. Serr `a, S. Pascual, J. Pons, R. O. Araz, and D. Scaini, “Universal speech enhancement with score-based diffusion,” arXiv:2206.03065, 2022

    2022 arXiv

  2. [9]

    Speech enhancement and dereverberation with diffusion-based generative models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerkmann, “Speech enhancement and dereverberation with diffusion-based generative models,”IEEE/ACM TASLP, vol. 31, pp. 2351–2364, 2023

    2023

  3. [10]

    Schr ¨odinger bridge for generative speech enhancement,

    A. Juki ´c, R. Korostik, J. Balam, and B. Ginsburg, “Schr ¨odinger bridge for generative speech enhancement,” inInterspeech, 2024, pp. 1175–1179

    2024

  4. [11]

    Diffusion-based speech enhancement with schr\

    S. Wang, S. Liu, A. Harper, P. Kendrick, M. Salz- mann, and M. Cernak, “Diffusion-based speech enhancement with schr\” odinger bridge and symmetric noise schedule,” arXiv:2409.05116, 2024

    2024 arXiv

  5. [12]

    Investigating training objectives for generative speech enhancement,

    J. Richter, D. De Oliveira, and T. Gerkmann, “Investigating training objectives for generative speech enhancement,” inIEEE ICASSP, 2025, pp. 1–5

    2025

  6. [13]

    Gan-based speech enhancement for low snr using latent feature conditioning,

    S. S. Shetu, E. A. P. Habets, and A. Brendel, “Gan-based speech enhancement for low snr using latent feature conditioning,” in IEEE ICASSP, 2025, pp. 1–5

    2025

  7. [14]

    On the behavior of intrusive and non-intrusive speech enhancement metrics in predictive and generative settings,

    D. de Oliveira, J. Richter, J.-M. Lemercier, T. Peer, and T. Gerk- mann, “On the behavior of intrusive and non-intrusive speech enhancement metrics in predictive and generative settings,” in Speech Communication; 15th ITG Conference, 2023, pp. 260– 264

    2023

  8. [15]

    Score-based generative modeling through stochas- tic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochas- tic differential equations,” inICLR, 2021

    2021

  9. [16]

    Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m ´ecanique quantique,

    E. Schr ¨odinger, “Sur la th ´eorie relativiste de l’ ´electron et l’interpr´etation de la m ´ecanique quantique,” inAnnales de l’institut Henri Poincar´e, vol. 2, no. 4, 1932, pp. 269–310

    1932

  10. [18]

    Likelihood training of schr ¨odinger bridge using forward-backward sdes theory,

    T. Chen, G.-H. Liu, and E. A. Theodorou, “Likelihood training of schr ¨odinger bridge using forward-backward sdes theory,” in ICLR, 2022

    2022

  11. [19]

    Diffusion schr¨odinger bridge matching,

    Y . Shi, V . De Bortoli, A. Campbell, and A. Doucet, “Diffusion schr¨odinger bridge matching,” inNeurIPS, vol. 36, 2023, pp. 62 183–62 223

    2023

  12. [20]

    I2 sb: Image-to-image schr ¨odinger bridge,

    G.-H. Liu, A. Vahdat, D.-A. Huang, E. A. Theodorou, W. Nie, and A. Anandkumar, “I2 sb: Image-to-image schr ¨odinger bridge,” in ICML, 2023

    2023

  13. [21]

    Schrodinger bridges beat diffusion models on text-to-speech synthesis,

    Z. Chen, G. He, K. Zheng, X. Tan, and J. Zhu, “Schrodinger bridges beat diffusion models on text-to-speech synthesis,” arXiv:2312.03491, 2023

    2023 arXiv

  14. [22]

    Bridge-sr: Schr¨odinger bridge for efficient sr,

    C. Li, Z. Chen, F. Bao, and J. Zhu, “Bridge-sr: Schr¨odinger bridge for efficient sr,” inIEEE ICASSP, 2025, pp. 1–5

    2025

  15. [23]

    Diffusion-based speech enhance- ment: Demonstration of performance and generalization,

    J. Richter and T. Gerkmann, “Diffusion-based speech enhance- ment: Demonstration of performance and generalization,”Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation

    2024

  16. [24]

    Tackling the generative learn- ing trilemma with denoising diffusion GANs,

    Z. Xiao, K. Kreis, and A. Vahdat, “Tackling the generative learn- ing trilemma with denoising diffusion GANs,” inICLR, 2022

    2022

  17. [25]

    Semi-implicit denoising diffusion models (sid- dms),

    y. xu, M. Gong, S. Xie, W. Wei, M. Grundmann, K. Batmanghe- lich, and T. Hou, “Semi-implicit denoising diffusion models (sid- dms),” inNeurIPS, vol. 36, 2023, pp. 17 383–17 394

    2023

  18. [26]

    Ufogen: You forward once large scale text-to-image generation via diffusion gans,

    Y . Xu, Y . Zhao, Z. Xiao, and T. Hou, “Ufogen: You forward once large scale text-to-image generation via diffusion gans,” in IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2024, pp. 8196–8206

    2024

  19. [27]

    Ears: An anechoic fullband speech dataset benchmarked for speech enhancement and derever- beration,

    J. Richter, Y .-C. Wu, S. Krenn, S. Welker, B. Lay, S. Watan- abe, A. Richard, and T. Gerkmann, “Ears: An anechoic fullband speech dataset benchmarked for speech enhancement and derever- beration,” inInterspeech, 2024, pp. 4873–4877

    2024

  20. [28]

    Adversarial schr ¨odinger bridge matching,

    N. Gushchin, D. Selikhanovych, S. Kholkin, E. Burnaev, and A. Korotin, “Adversarial schr ¨odinger bridge matching,” in NeurIPS, vol. 37, 2024, pp. 89 612–89 651

    2024

  21. [29]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”TMLR, 2023

    2023

  22. [30]

    Diffusion-based generative speech source separation,

    R. Scheibler, Y . Ji, S.-W. Chung, J. Byun, S. Choe, and M.-S. Choi, “Diffusion-based generative speech source separation,” in IEEE ICASSP, 2023, pp. 1–5

    2023

  23. [31]

    High-fidelity audio compression with improved rvqgan,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Ku- mar, “High-fidelity audio compression with improved rvqgan,” in NeurIPS, vol. 36, 2023, pp. 27 980–27 993

    2023

  24. [32]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inICLR, 2019

    2019

  25. [33]

    Improved techniques for training score- based generative models,

    Y . Song and S. Ermon, “Improved techniques for training score- based generative models,” inNeurIPS, vol. 33, 2020, pp. 12 438– 12 448

    2020

  26. [34]

    Perceptual eval- uation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,

    A. Rix, J. Beerends, M. Hollier, and A. Hekstra, “Perceptual eval- uation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” inIEEE Interna- tional Conference on Acoustics, Speech, and Signal Processing, vol. 2, 2001, pp. 749...

    2001

  27. [35]

    Sdr – half-baked or well done?

    J. L. Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “Sdr – half-baked or well done?” inIEEE ICASSP, 2019, pp. 626–630

    2019

  28. [36]

    An algorithm for predicting the intelligi- bility of speech masked by modulated noise maskers,

    J. Jensen and C. H. Taal, “An algorithm for predicting the intelligi- bility of speech masked by modulated noise maskers,”IEEE/ACM TASLP, vol. 24, no. 11, pp. 2009–2022, 2016

    2009

  29. [37]

    Dnsmos: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. A. Reddy, V . Gopal, and R. Cutler, “Dnsmos: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” inIEEE ICASSP, 2021, pp. 6493–6497

    2021

  30. [38]

    Icassp 2024 speech signal improvement challenge,

    N.-C. Ristea, A. Saabas, R. Cutler, B. Naderi, S. Braun, and S. Branets, “Icassp 2024 speech signal improvement challenge,” inIEEE International Conference on Acoustics, Speech, and Sig- nal Processing Workshops, 2024, pp. 15–16

    2024

  31. [39]

    Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,

    Y . Luo and N. Mesgarani, “Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,”IEEE/ACM TASLP, vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  32. [40]

    Hybrid transformers for music source separation,

    S. Rouard, F. Massa, and A. D ´efossez, “Hybrid transformers for music source separation,” inIEEE ICASSP, 2023, pp. 1–5

    2023

  33. [41]

    torch-pesq: Loss function inspired by the pesq score,

    I. A. L. Erlangen, “torch-pesq: Loss function inspired by the pesq score,” https://github.com/audiolabs/torch-pesq, 2023, ac- cessed: 2025-01-03

    2023

  34. [42]

    The pesqetarian: On the relevance of goodhart’s law for speech en- hancement,

    D. de Oliveira, S. Welker, J. Richter, and T. Gerkmann, “The pesqetarian: On the relevance of goodhart’s law for speech en- hancement,” inInterspeech, 2024, pp. 3854–3858

    2024

  35. [43]

    Nu-wave: A diffusion probabilistic model for neural audio upsampling,

    J. Lee and S. Han, “Nu-wave: A diffusion probabilistic model for neural audio upsampling,” inInterspeech, 2021, pp. 1634–1638

    2021

Pith tools

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