Pith. sign in

REVIEW 4 major objections 5 minor 44 references

The paper claims that training a Mamba backbone with a Schrödinger Bridge objective collapses the optimal transport path from degraded to clean speech into a single inference step, outperforming generative baselines that need 10-50 steps wh

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A Mamba-based speech enhancer trained with Schrödinger Bridge objectives produces strong denoising and dereverberation in one inference step with a low real-time factor.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful SB+Mamba combination for one-step speech enhancement, but the published claims run ahead of the evidence. the 4 major comments →

arxiv 2510.16834 v3 pith:3VG7JBJ5 submitted 2025-10-19 cs.SD cs.AIcs.LGeess.AS

Schr\"odinger Bridge Mamba for One-Step Speech Enhancement

Jing Yang , Sirui Wang , Chao Wu , Lei Guo , Fan Fan This is my paper
classification cs.SD cs.AIcs.LGeess.AS
keywords Schrödinger BridgeMambaspeech enhancementone-step inferencegenerative speech enhancementdenoisingdereverberationstate-space models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 proposes Schrödinger Bridge Mamba (SBM), a speech enhancement model that trains a Mamba state-space backbone under a Schrödinger Bridge objective. The central claim is that Mamba's Markovian dynamics and the Schrödinger Bridge's stochastic process align, so a single forward pass at the start of the reverse process maps degraded speech to clean speech without iterative refinement. In a joint denoising and dereverberation task spanning four test sets, SBM with one inference step outperforms or matches Schrödinger Bridge models that require 10-50 steps, as well as one-step Schrödinger Bridge variants and a predictive-mapping Mamba baseline, while achieving the smallest real-time factor. If this holds, generative speech enhancement can be high-quality and real-time even on lightweight devices.

Core claim

SBM's discovery is that the Schrödinger Bridge training paradigm, previously tied to heavy score-network backbones and iterative sampling, can be paired with a selective state-space backbone so that the optimal transport between degraded and clean speech is collapsed into a one-step state transition. The paper argues this works because Mamba's hidden-state evolution is Markovian, matching the SB stochastic process, and its selective mechanism can dynamically parameterize the optimal control along the path. Empirically, the one-step SBM outperforms baselines on perceptual, semantic, and speaker-similarity metrics, with a real-time factor of 0.0048 and only 3.93M parameters.

What carries the argument

The key machinery is the SBM training-inference scheme: the Schrödinger Bridge objective trains a Mamba backbone using data-prediction losses on interpolated states, with a timestep embedding added to the Mamba blocks; inference then runs the backbone once at timestep T=1. The compatibility between SB and Mamba is the Markov property shared by the SB reverse process and Mamba's state recurrence h_t = A h_{t-1} + B u_t, which lets the model act as an implicit one-step optimal transport map.

Load-bearing premise

The entire efficiency and quality claim rests on the assumption that one forward pass of the Mamba backbone at timestep T=1 faithfully reproduces the full multi-step Schrödinger Bridge reverse transport, yet the paper offers no proof, error bound, or distillation argument for that jump.

What would settle it

Run the trained SBM in iterative mode (e.g., taking multiple small reverse steps with the same network) and compare output quality to the one-step output; if iterative inference clearly outperforms one-step, the one-step claim is falsified. Alternatively, evaluate one-step SBM on noise conditions with SNRs far outside the training range; a sharp quality collapse would indicate the model approximates a mapping rather than the true transport.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • One-step inference makes generative speech enhancement practical for real-time and edge applications, with a real-time factor below 0.005.
  • Switching from predictive mapping to SB training improves the same Mamba backbone on all test sets, with no increase in inference cost.
  • The same SB+Mamba concept could be applied to other generative restoration tasks such as super-resolution and inpainting.
  • Because SBM inherits Mamba's streaming and linear-complexity properties, it can be further optimized for streaming devices.
  • The method closes the performance gap between one-step generative enhancement and slower iterative models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The one-step jump from T=1 is an empirical shortcut; the paper gives no error bound, so the claim that a single forward pass faithfully approximates the Schrödinger Bridge transport rests entirely on experimental evidence.
  • The Markov-property compatibility argument is suggestive but not a formal equivalence proof; a careful study of how Mamba's discretized SSM relates to the SB SDE could turn the architectural alignment into a provable guarantee.
  • If SBM is extended to other tasks, the same training recipe (SB objective + timestep embedding + one-step inference) is a testable baseline for image, video, and cross-modal generation.
  • Direct ablations that isolate the architecture from the training paradigm (e.g., same Mamba backbone trained with predictive mapping vs. SB, and same SB objective with different backbones) would further support the claimed synergy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces Schrödinger Bridge Mamba (SBM), a speech-enhancement model that trains a Mamba-based backbone with a Schrödinger Bridge (SB) training objective and performs inference in a single forward pass at timestep T=1. The authors argue that Mamba's Markovian state evolution is naturally compatible with the SB framework, and they evaluate SBM on joint denoising and dereverberation using DNS and VoiceBank-Demand testsets. They report that 1-step SBM outperforms or matches several baselines, including SB-NCSN++, SBCTM, SB-UFOGen, and a predictive-mapping Mamba model (Mamba-base), while achieving the lowest real-time factor in Table 3.

Significance. If the empirical claims hold, the paper makes a useful contribution by replacing the heavy NCSN++ backbone used in most SB-based speech enhancement methods with a much lighter Mamba backbone, thereby enabling one-step inference without sacrificing quality. The controlled comparison against Mamba-base (same backbone, same data, different training paradigm) is the strongest part of the evidence and suggests that the SB training objective itself provides a gain over predictive mapping. The paper also articulates an interesting, though heuristic, conceptual link between SB and selective state-space models. However, the headline claim of 'outperforming strong baselines' is weakened by the uncontrolled SBCTM baseline, the absence of the architecture ablations promised in the abstract, lack of error bars on small metric margins, and an unexamined one-step-at-T=1 approximation. These issues are fixable and do not invalidate the core idea, but they need to be addressed before the paper can be accepted.

major comments (4)
  1. [Section 4.3, Tables 1-2] The SBCTM baseline is not retrained on the same data. Section 4.3 states that SBCTM's open-sourced pretrained model was used directly, while all other baselines were retrained on the SBM training corpus. Since SBCTM is one of the two one-step SB variants in the comparison, its low scores on DNS With Reverb (e.g., PESQ 1.218 and SpeechBERTScore 0.494 vs SBM 1.971 and 0.784) could be due to mismatched training data, noise/reverberation simulation, or SNR schedule rather than to SBM's superiority. Conversely, SBCTM beats SBM on DNS No Reverb OVRL (3.298 vs 3.292) and PESQ (2.835 vs 2.825), and on VoiceBank PESQ (3.558 vs 3.503). The central comparative claim therefore requires either retraining SBCTM under the same conditions or explicitly restricting the claim to models trained on the same data.
  2. [Abstract vs. Section 5] The abstract promises that 'Ablation studies reveal that the SB paradigm consistently yields improved performance across diverse architectures' and that 'Mamba exhibits a stronger performance under the SB paradigm compared to Multi-Head Self-Attention (MHSA) and Long Short-Term Memory (LSTM) backbones.' The body contains no such ablation table or experiment. Section 5 only compares SBM with Mamba-base, SB-NCSN++, SBCTM, and SB-UFOGen. The claimed MHSA/LSTM comparison is absent, and the 'diverse architectures' claim is unsupported. This missing evidence is load-bearing for the paper's conceptual contribution, not merely a presentation detail.
  3. [Tables 1-2] No error bars, confidence intervals, or significance tests are reported. Several of the key comparisons involve very small differences, e.g., DNS No Reverb OVRL (SBM 3.292 vs SBCTM 3.298), DNS No Reverb SpeechBERTScore (SBM 0.893 vs Mamba-base 0.887), and VoiceBank PESQ (SBM 3.503 vs SBCTM 3.558, where SBM is worse). Without variance estimates or paired tests, the claim that SBM 'outperforms' these baselines on those metrics is not statistically supported, especially on testsets of only 150–824 utterances. The authors should report per-seed variance, confidence intervals, or at least a paired test on the held-out sets.
  4. [Section 3 and Table 3] The one-step inference design at timestep T=1 is central to the efficiency claim, but it is not analyzed or ablated. Section 3 simply states that 'The inference process is only one step with timestep T=1 representing the start of reverse SDE as in the typical iterative sampling process.' No error bound, distillation argument, or convergence analysis is provided, and no comparison to e.g. T=2, T=5, or T=10 is shown. Without such evidence, it is unclear whether the performance at T=1 reflects a genuine property of SB-trained Mamba or a particular choice of loss and timestep. At minimum, an ablation over inference timesteps would clarify the trade-off and support the 'one-step' claim.
minor comments (5)
  1. [Section 4 heading] The heading 'EXPERIEMENTAL SETUP' contains a typo; should be 'EXPERIMENTAL SETUP'.
  2. [Table 3] RTF measurement details are insufficient: no GPU model, no batch size, no input duration beyond '10 pieces of 10s audio', and no standard deviation across the 10 pieces. This makes the RTF comparison difficult to reproduce.
  3. [Section 2.2] Equation (2) presents a linear state-space recurrence, but Mamba's selectivity is usually introduced by making B and C (and sometimes A) input-dependent. The authors mention this informally, but the notation may mislead readers; a brief clarification would help.
  4. [Section 4.3] For the retrained baselines, no details about hyperparameters, training iterations, or validation-based checkpoint selection are given. This makes it hard to assess whether the baselines were tuned to a comparable degree.
  5. [Section 5] The statement 'SBM outperforms all baselines in all metrics on the DNS Real Recordings' is supported by Table 2, but the broader claim that SBM 'outperforms strong baselines' is too strong given the exceptions on DNS No Reverb and VoiceBank. The wording should be softened to 'outperforms or matches' where appropriate.

Circularity Check

0 steps flagged

No significant circularity: SBM's contributions are empirical and validated against external benchmarks.

full rationale

No circular derivation chain is present. SBM's novelty is to train a Mamba backbone with a Schrödinger Bridge objective and then use one-step inference; this is an architectural/training choice whose reported quality is measured on external test sets, not a quantity that is defined in terms of the outcome. The SB–Mamba compatibility discussion in Section 3 is explicitly speculative ('A natural question thus arises... A plausible conjecture...') and is tested experimentally rather than used to derive the scores. The comparison with Mamba-base uses the same backbone and same training data, providing a controlled check that the SB objective, not a backbone change, drives the improvement. The decision to use SBCTM's open-sourced pretrained model without retraining (Section 4.3: 'SBCTM open-sourced pretrained model and we used it in our experiments. For the other baselines without open-sourced models, ... we re-trained models ... using the same datasets as we used to train SBM') is a potential fairness confound for the 'outperforms' claim, but it is an experimental-control concern, not circularity: SBM's numbers are not constructed from SBCTM's outputs. The self-citations to AnyEnhance-related work ([35], [38]) are only for metric pretrained models and an external SOTA comparison, not load-bearing derivation or uniqueness. The one-step T=1 inference lacks a formal error bound, but this is an unsupported assumption about approximation quality, not a case where the prediction is equivalent to the input by construction. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central results rest on three imported/asserted building blocks: (1) the Gaussian SB state parameterization from prior work, (2) the heuristic Markov/control alignment between SB and Mamba, and (3) the unproven one-step jump at T=1. The loss weights and VE noise-schedule constants are free parameters the paper does not report. No new entities are introduced.

free parameters (4)
  • VE noise schedule parameters (sigma_min, sigma_max, schedule shape) = not reported; chosen via pre-experiments over VP and Bridge-gmax
    The SB state parameterization x_t = mu_x(t) + sigma_x(t) z depends on the chosen noise schedule; Section 4.1 says VE was used due to better performance in pre-experiments, but the schedule constants are not given.
  • loss weights lambda_1..lambda_4 = not reported
    The SB data-prediction loss L = lambda_1 L_mse(S,S_hat) + lambda_2 L_mse(|S|,|S_hat|) + lambda_3 L_mr,mse(S,S_hat) + lambda_4 L_mr,mse(|S|,|S_hat|) is defined in Section 4.1 without numerical values; these weights affect the trained model and are not justified.
  • timestep T=1 for one-step inference = 1
    Inference is fixed to a single forward pass at t=T=1 (Section 3); this is a design choice, not derived from the SB SDE solution.
  • Mamba backbone hyperparameters (width, depth, state dimension) = not reported
    The backbone follows oSpatialNet [16] with a full-band block [24], but no architecture hyperparameters or training schedule are given, leaving key configuration choices unspecified.
axioms (5)
  • domain assumption The SB transport between degraded and clean distributions has Gaussian marginals x_t = mu_x(t) + sigma_x(t) z, mu_x = w_x x + w_y y (from Jukic et al.)
    Invoked in Section 2.1 to define the training states; it imports the Gaussian SB solution from prior work without re-deriving it here.
  • standard math Schrödinger Bridge is a Markov stochastic process satisfying boundary conditions (Chen et al. 2021)
    Section 2.1 depends on this characterization for the forward/reverse SDE pair.
  • ad hoc to paper Mamba's recurrence h_t = A h_(t-1) + B u_t is a Markov process and can therefore host SB-style trajectories
    Sections 2.2 and 3 use the Markov property of the hidden state to argue compatibility with SB; A and B are input-dependent in Mamba, and the paper gives no theorem that this state process matches an SB path measure.
  • ad hoc to paper A single evaluation at timestep T=1 with the data-prediction objective suffices to approximate the full SB reverse transport
    Section 3: inference is only one step with timestep T=1 representing the start of reverse SDE; no convergence or error bound is provided for this one-step jump, so the central efficiency claim depends on it.
  • ad hoc to paper Training a Mamba model resembles learning an optimal control strategy that aligns with SB stochastic control
    Section 3 asserts this alignment as a plausible conjecture; it motivates the method but is not a proven equivalence.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Schr\"odinger Bridge Mamba for One-Step Speech Enhancement." pith.science (2026). https://pith.science/paper/3VG7JBJ5

@misc{pith2026251016834,
  author       = {Pith},
  title        = {Pith review of: Schr\"odinger Bridge Mamba for One-Step Speech Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3VG7JBJ5}},
  note         = {Machine review of arXiv:2510.16834}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present Schr\"odinger Bridge Mamba (SBM), a novel model for efficient speech enhancement by integrating the Schr\"odinger Bridge (SB) training paradigm and the Mamba architecture. Experiments of joint denoising and dereverberation tasks demonstrate SBM outperforms strong generative and discriminative methods on multiple metrics with only one step of inference while achieving a competitive real-time factor for streaming feasibility. Ablation studies reveal that the SB paradigm consistently yields improved performance across diverse architectures over conventional mapping. Furthermore, Mamba exhibits a stronger performance under the SB paradigm compared to Multi-Head Self-Attention (MHSA) and Long Short-Term Memory (LSTM) backbones. These findings highlight the synergy between the Mamba architecture and the SB trajectory-based training, providing a high-quality solution for real-world speech enhancement. Demo page: https://sbmse.github.io

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 15 linked inside Pith

  1. [1]

    INTRODUCTION Deep generative models have been increasingly employed for speech enhancement (SE) tasks. By learning the underlying distribution of clean audio given its degraded counterpart, generative models are capable of generating high-quality speech from low-quality in- puts that include noise, reverberation, clipping, bandwidth limita- tion or a mixt...

  2. [2]

    PRELIMINARIES 2.1. Schr ¨odinger Bridge for Speech Enhancement Schr¨odinger Bridge (SB) seeks the path measureQthat optimally interpolates between two arbitrary boundary distributionsx 0 ∼p 0 arXiv:2510.16834v1 [cs.SD] 19 Oct 2025 Huawei Proprietary - Restricted Distribution2 Noise schedule State 𝑥𝑥𝑡𝑡 Timestep 𝑡𝑡 time embedding NCSN++ backboneClean data 𝑥...

    Pith/arXiv arXiv 2025

  3. [3]

    Previous sections have elucidated the alignment of SB and Mamba from the Markov process perspective

    SCHR ¨ODINGER BRIDGE MAMBA This section will first discuss the connections between SB and Mamba, and then present our implementation ofSchr ¨odinger Bridge Mamba (SBM)for generative speech enhancement. Previous sections have elucidated the alignment of SB and Mamba from the Markov process perspective. From the perspective of controlled dynamical system, t...

  4. [4]

    Implementation Details To cover common real-life scenarios, we focus on a joint denois- ing and dereverberation task for speech enhancement

    EXPERIEMENTAL SETUP 4.1. Implementation Details To cover common real-life scenarios, we focus on a joint denois- ing and dereverberation task for speech enhancement. To im- plement SBM, we follow the SB training paradigm in [2] with VE noise schedule due to its better performance than VP [2] and Bridge-gmax noise schedules [7] in our pre-experiments. We i...

  5. [5]

    RESULTS AND DISCUSSION Table 1 and Table 2 demonstrate superior performance of SBM across all four benchmark testsets. SBM outperforms all baselines in all metrics on the DNS Real Recordings that represent degraded audio collected in the wild, indicating the strong capability of SBM to deal with real-world enhancement tasks. On the DNS With Reverb, SBM ac...

  6. [6]

    CONCLUSION We introduce Schr¨odinger Bridge Mamba (SBM), a new concept of training-inference framework in deep generative models motivated by the inherent compatibility between Schr ¨odinger Bridge (SB) paradigm and selective state-space model Mamba. We present a SBM implementation for a popular, fundamental audio processing task – speech enhancement, and...

  7. [7]

    Speech enhancement and dereverberation with diffusion-based generative models,

    J. Richter, S. Welker, J.-M. Lemercier, B. Lay, and T. Gerk- mann, “Speech enhancement and dereverberation with diffusion-based generative models,”TASLP, 2023

    2023

  8. [8]

    Schr¨odinger bridge for generative speech enhancement,

    A. Juki ´c, R. Korostik, J. Balam, and B. Ginsburg, “Schr¨odinger bridge for generative speech enhancement,” arXiv:2407.16074, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Diffusion-based speech enhancement with schr ¨odinger bridge and symmetric noise schedule,

    S. Wang, S. Liu, A. Harper, et al., “Diffusion-based speech enhancement with schr ¨odinger bridge and symmetric noise schedule,”arXiv:2409.05116, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Investigating training objectives for generative speech enhancement,

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

    2025

  11. [11]

    Few-step adversarial schr¨odinger bridge for generative speech enhancement,

    S. Han, S. Lee, J. Lee, and K. Lee, “Few-step adversarial schr¨odinger bridge for generative speech enhancement,”arXiv e-prints, 2025

    2025

  12. [12]

    Schr ¨odinger bridge consistency trajectory models for speech enhancement,

    S. Nishigori, K. Saito, N. Murata, et al., “Schr ¨odinger bridge consistency trajectory models for speech enhancement,”arXiv e-prints, 2025

    2025

  13. [13]

    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,” inICASSP. IEEE, 2025, pp. 1–5

    2025

  14. [14]

    A2sb: Audio-to-audio schrodinger bridges,

    Z. Kong, K.J Shih, W. Nie, et al., “A2sb: Audio-to-audio schrodinger bridges,”arXiv:2501.11311, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    I2sb: Image-to- image schr¨odinger bridge,

    G.H. Liu, A. Vahdat, D.A. Huang, et al., “I2sb: Image-to- image schr¨odinger bridge,”arXiv:2302.05872, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    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

    Pith/arXiv arXiv 2023

  17. [17]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D.P. Kingma, A. Kumar, S. Er- mon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inICLR

  18. [18]

    Score-based generative modeling in latent space,

    A. Vahdat, K. Kreis, and J. Kautz, “Score-based generative modeling in latent space,”NeurIPS, 2021

    2021

  19. [19]

    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,” inCVPR, 2024, pp. 8196–8206

    2024

  20. [20]

    Consistency trajectory models: Learning probability flow ode trajectory of diffusion,

    D. Kim, C.-H. Lai, W.-H. Liao, et al., “Consistency trajectory models: Learning probability flow ode trajectory of diffusion,” arXiv preprint arXiv:2310.02279, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Multichannel long-term streaming neural speech enhancement for static and moving speakers,

    C. Quan and X. Li, “Multichannel long-term streaming neural speech enhancement for static and moving speakers,”IEEE Signal Processing Letters, vol. 31, pp. 2295–2299, 2024

    2024

  23. [23]

    An investiga- tion of incorporating mamba for speech enhancement,

    R. Chao, W.H. Cheng, M. La Quatra, et al., “An investiga- tion of incorporating mamba for speech enhancement,” inSLT. IEEE, 2024

    2024

  24. [24]

    Universal speech enhancement with regression and generative mamba,

    R. Chao, R. Nasretdinov, Y .C.F. Wang, A. Juki´c, S.W. Fu, and Y . Tsao, “Universal speech enhancement with regression and generative mamba,”arXiv:2505.21198, 2025

    arXiv 2025

  25. [25]

    Stochastic con- trol liaisons: Richard sinkhorn meets gaspard monge on a schrodinger bridge,

    Y . Chen, T.T Georgiou, and M. Pavon, “Stochastic con- trol liaisons: Richard sinkhorn meets gaspard monge on a schrodinger bridge,”Siam Review, 2021

    2021

  26. [26]

    A new approach to linear filtering and predic- tion problems,

    R.E. Kalman, “A new approach to linear filtering and predic- tion problems,” 1960

    1960

  27. [27]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,”arXiv:2111.00396, 2021

    Pith/arXiv arXiv 2021

  28. [28]

    From s4 to mamba: A comprehensive survey on structured state space models,

    S. Somvanshi, M.M. Islam, M.S. Mimi, et al., “From s4 to mamba: A comprehensive survey on structured state space models,”arXiv:2503.18970, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Ultra dual-path compression for joint echo cancellation and noise suppression,

    H. Chen, J. Yu, Y . Luo, R. Gu, W. Li, Z. Lu, and C. Weng, “Ultra dual-path compression for joint echo cancellation and noise suppression,”arXiv:2308.11053, 2023

    Pith/arXiv arXiv 2023

  30. [30]

    Spatialnet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,

    C. Quan and X. Li, “Spatialnet: Extensively learning spatial in- formation for multichannel joint speech separation, denoising and dereverberation,”TASLP, vol. 32, pp. 1310–1323, 2024

    2024

  31. [31]

    The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

    C.KA Reddy, V . Gopal, R. Cutler, et al., “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,”arXiv:2005.13981, 2020

    Pith/arXiv arXiv 2020

  32. [32]

    Aishell- 3: A multi-speaker mandarin tts corpus and the baselines,

    Y . Shi, H. Bu, X. Xu, S. Zhang, and M. Li, “Aishell- 3: A multi-speaker mandarin tts corpus and the baselines,” arXiv:2010.11567, 2020

    Pith/arXiv arXiv 2010

  33. [33]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” inICASSP. IEEE, 2015, pp. 5206–5210

    2015

  34. [34]

    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,”TASLP, vol. 30, pp. 829–852, 2021

    2021

  35. [35]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M.L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” inICASSP. IEEE, 2017, pp. 5220–5224

    2017

  36. [36]

    Pyroomacous- tics: A python package for audio room simulation and array processing algorithms,

    R. Scheibler, E. Bezzam, and I. Dokmani ´c, “Pyroomacous- tics: A python package for audio room simulation and array processing algorithms,” inICASSP. IEEE, 2018, pp. 351–355

    2018

  37. [37]

    In- vestigating rnn-based speech enhancement methods for noise- robust text-to-speech,

    C.V . Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “In- vestigating rnn-based speech enhancement methods for noise- robust text-to-speech,” inISCA, 2016, pp. 159–165

    2016

  38. [38]

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

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

    2021

  39. [39]

    Nisqa: A deep cnn-self-attention model for multidimen- sional speech quality prediction with crowdsourced datasets,

    G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “Nisqa: A deep cnn-self-attention model for multidimen- sional speech quality prediction with crowdsourced datasets,” arXiv:2104.09494, 2021

    Pith/arXiv arXiv 2021

  40. [40]

    Speechbertscore: Reference-aware automatic evaluation of speech generation leveraging nlp evaluation met- rics,

    T. Saeki, S. Maiti, S. Takamichi, S. Watanabe, and H. Saruwatari, “Speechbertscore: Reference-aware automatic evaluation of speech generation leveraging nlp evaluation met- rics,”arXiv:2401.16812, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Anyenhance: A unified generative model with prompt-guidance and self-critic for voice enhancement,

    J. Zhang, J. Yang, Z. Fang, Y . Wang, Z. Zhang, Z. Wang, F. Fan, and Z. Wu, “Anyenhance: A unified generative model with prompt-guidance and self-critic for voice enhancement,” arXiv:2501.15417, 2025

    arXiv 2025

  42. [42]

    Exten- sion of itu-t recommendation p. 862 pesq towards measuring speech intelligibility with vocoders,

    J.G. Beerends, S. Van Wijngaarde, and R. Van Buuren, “Exten- sion of itu-t recommendation p. 862 pesq towards measuring speech intelligibility with vocoders,” 2005

    2005

  43. [43]

    An algorithm for predicting the in- telligibility of speech masked by modulated noise maskers,

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

    2009

  44. [44]

    Multi-metric preference alignment for generative speech restoration,

    J. Zhang, X. Zhang, J. Yang, Y . Wang, F. Fan, and Z. Wu, “Multi-metric preference alignment for generative speech restoration,”arXiv:2508.17229, 2025

    arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.