Pith. sign in

REVIEW 3 major objections 6 minor 55 references

HybridSB-MoE: Dual-Domain Schr\"odinger Bridges with Scene-Adaptive Expert Routing for Speech Enhancement

T0 review · 3 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Speech enhancement routes between two error types to reach PESQ 3.88

desk verdict A novel hybrid architecture with strong reported numbers, but the central Theorem 1 is not proved as written and the small-K guarantee is currently a heuristic. read the letter →

arxiv 2608.12715 v1 pith:M3GKZ7SA submitted 2026-08-13 cs.SD cs.AI

Zhengyi Lu , Aswini Sivakumar , Jie Hu , Yao Qiang This is my paper · ORCID
classification cs.SDcs.AI
keywords speechenhancementSchrödingerbridgemixtureofexpertsuncertaintyfusiondiscretizationboundfew-stepinferencedual-domainpathconsistency
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 takes on the long-standing split between spectral and waveform speech enhancement and argues that the two should not be averaged together but routed between, depending on which kind of error is active. It builds a dual-domain system: a spectral mixture-of-experts path whose expert disagreement measures epistemic uncertainty, and a waveform Schrödinger bridge whose intrinsic noise measures aleatoric uncertainty. The central theoretical claim is that two training regularizers, path consistency and trajectory anchoring, together bound the eight-step bridge sampling error in Wasserstein distance, so the small sampling budget follows from the training objective rather than being a heuristic. On the VoiceBank+DEMAND benchmark the system reports a PESQ of 3.88, above diffusion and Schrödinger-bridge baselines and competitive with one-step consistency-distilled methods. If the design works as argued, dual-domain speech enhancement becomes a problem of selecting between error regimes instead of averaging predictions.

What carries the argument

The load-bearing identity is the discretization bound of Theorem 1, $W_2(\hat{p}_K, p^{\mathrm{br}}_0) \le C_1 K^{-\alpha} + C_2 \sqrt{L^*_{\mathrm{path}} + L^*_{\mathrm{traj}}}$, which decomposes the $K$-step sampling error into a schedule-controlled term and a training-controlled term. The schedule term $C_1 K^{-\alpha}$ with $\alpha=\min(1,\gamma)$ is driven by the front-loading exponent $\gamma$; the regularizer term is controlled by the path-consistency loss (cross-timestep agreement of clean predictions) and the trajectory loss (anchoring each state to the schedule-consistent reconstruction). The other load-bearing mechanism is the asymmetric fusion: the spectral MoE's top-$k$ expert disagreement $u_{\mathrm{epi}}$ is treated as epistemic uncertainty and the bridge's learned log-variance head $u_{\mathrm{ale}}$ as aleatoric uncertainty, and a two-layer MLP maps their z-scored values to a scalar weight that picks between the two reconstruction paths. These are coupled: the MoE produces the error $u_{\mathrm{epi}}$ detects, the regularized SB produces the error $u_{\mathrm{ale}}$ detects, and Theorem 1 controls the SB-side error at the deployment budget.

What would settle it

Estimate the Lipschitz constant of the trained U-Net denoiser on a batch of bridge states and plot the observed PESQ or Wasserstein gap against the number of sampling steps. If the curve saturates far earlier than the predicted $C_1K^{-\alpha}$ term with the estimated constant, or if retraining without the two regularizers still gives high PESQ at $K=8$, then Theorem 1 is not the mechanism that makes few-step inference work.

Watch

Extended reading notes

Core claim

The paper's central claim is that a speech enhancement system can profitably couple a spectral path and a waveform path if each carries a categorically different uncertainty signal, and that the waveform path's few-step inference cost can be tied to its training objective. The spectral path consists of five architecturally distinct experts with top-$k=2$ routing; disagreement among the selected experts yields an epistemic signal that indicates which inductive bias is failing. The waveform path is a Schrödinger bridge trained with path-consistency and trajectory regularizers; Theorem 1 states that with a front-loaded schedule $t_k=T(k/K)^\gamma$, the $2$-Wasserstein distance between the $K$-step rollout law and the continuous bridge marginal is bounded by $C_1 K^{-\alpha} + C_2 \sqrt{L^*_{\mathrm{path}} + L^*_{\mathrm{traj}}}$, $\alpha=\min(1,\gamma)$. This makes $K=8$ a consequence of minimizing the regularizers. An asymmetric fusion weight $w=\sigma(\mathrm{MLP}(\tilde{u}_{\mathrm{epi}},\tilde{u}_{\mathrm{ale}}))$ selects between the two pathways: high expert disagreement defers to the waveform branch and high bridge variance defers to the spectral branch. On VoiceBank+DEMAND the system reports PESQ 3.88, CBAK 3.85, and COVL 4.82.

Load-bearing premise

The argument assumes, without enforcing it during training, that the denoising network changes only a bounded amount when its input or the timestep changes; if that bound is much larger than expected, the guarantee about eight-step inference gives no practical control.

Editorial extensions

If this is right

  • $K=8$ inference is tied to the training objective: if the two regularizers reach small terminal values, the schedule term saturates at modest $K$, so the system should not need diffusion-scale step counts.
  • The fusion weight is interpretable: high spectral expert disagreement pushes the output toward the waveform reconstruction, and high bridge variance pushes it toward the spectral reconstruction, rather than averaging two predictions.
  • On VoiceBank+DEMAND the system reports the top PESQ (3.88), CBAK (3.85), and COVL (4.82) among the compared diffusion-, bridge-, and consistency-distilled baselines under a unified evaluation protocol.
  • The calibration loss yields a fusion-weight ECE of 0.042, an order-of-magnitude reduction over an uncalibrated single-pathway baseline, making the uncertainty signals usable downstream.
  • The method runs at real-time factor 0.28 with 8 steps, giving a 4–5x speedup over diffusion- and bridge-based baselines at higher PESQ, and defines the quality-efficiency frontier together with a one-step consistency-distilled method.

Reading between the lines

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

  • I infer that the bound's practical force depends on untested Lipschitz constants: the paper assumes $L_x$ and $L_t$ in Assumption 1(i) but does not enforce or measure them, so the sharpest reading of Theorem 1 is as a design-justifying inequality rather than a tight predictor of the PESQ-versus-$K$ curve.
  • I infer that the same asymmetric-fusion design could transfer to other restoration tasks with two structurally different reconstructions, such as image deblurring with a frequency-domain and a spatial-domain branch, wherever one branch can produce disagreement and the other intrinsic stochasticity.
  • I infer that a direct test of the rate $K^{-\alpha}$ is feasible: sweeping $K$ and fitting the empirical Wasserstein or PESQ gap against $C_1 K^{-\alpha}$ would confirm or refute whether the schedule term is the actual driver at small $K$.
  • I infer that because the archetype set generalizes by adding new expert architectures per noise family, the scene-adaptivity claim would be stronger if tested on corpora with larger train/test noise distribution shift.
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

3 major / 6 minor

Summary. The paper proposes HybridSB-MoE, a dual-domain speech enhancement system that combines a heterogeneous mixture-of-experts spectral path with a Schrödinger-bridge waveform path, and fuses them through an asymmetric uncertainty mechanism that distinguishes epistemic disagreement in the spectral MoE from aleatoric variance in the waveform bridge. The central theoretical contribution is Theorem 1, which claims a K-step bridge-discretization bound in 2-Wasserstein distance in terms of the training-time path-consistency and trajectory regularizers, and is used to justify the K=8 inference budget as an objective-level guarantee rather than an empirical heuristic. The empirical section reports state-of-the-art PESQ of 3.88 on VoiceBank+DEMAND, along with ablations supporting the design and a calibration analysis for the fusion weights.

Significance. If the theoretical claim were fully established, the paper would make a useful conceptual contribution: linking the inference-time step budget to explicit training-time regularizers, rather than treating few-step sampling as a purely empirical choice. The empirical work is also credible in its breadth: unified baseline reproduction, ablations that isolate the three design components, scene-stratified results, and calibration measurements. The main weakness is that the central guarantee is not established as written: the proof has an endpoint gap at the first reverse step, and the bound's constants and residual terms are not estimated. These issues do not invalidate the empirical system, but they do undermine the paper's headline theoretical claim, so the manuscript needs substantive revision before the guarantee can be accepted.

major comments (3)
  1. [Appendix C.2/C.5, Eq. (8)] At t=T, Eq. (4) gives beta_T=0 and sigma_T=0, so the trajectory residual in Eq. (8) is identically zero for any value of the denoiser at (y,y,T). The path-consistency loss and data-prediction loss are expectations over continuous t, so the single point t=T has measure zero and does not constrain that value. Yet the K-step rollout of Eq. (5) initializes at x_T=y and evaluates the denoiser at exactly (y,y,T) in its first reverse step. Lemma 1's proof restricts Eq. (13) to [epsilon0,T-epsilon0], and the statement in Appendix C.5 that 'all sampling steps lie strictly in the interior' is false for the first update. Thus, as proved, Theorem 1 does not control the first reverse step, and the claimed objective-level justification for K=8 is not established. A fix would require either an explicit endpoint loss term, a rollout that starts at T-epsilon, or a separate Lipschitz-based argument covering t=T.
  2. [Appendix C.4] The synchronous coupling argument bounds W2 by accumulating per-step one-step errors delta_k, but Lemma 1 bounds those errors on true bridge states, whereas the rollout feeds the denoiser its own previous outputs. The proof asserts rather than demonstrates that the propagated states remain close enough for Lemma 1 to apply. The exp(L_x T) factor is absorbed into the unestimated constants C1 and C2, so the proof does not quantitatively control the accumulation of model error along the trajectory. Please provide a complete induction bounding the distance between the true bridge state and the rollout state at each step, or state the needed closeness condition as an explicit assumption.
  3. [Section 3, Theorem 1, Appendix C.6] The second term C2 sqrt(L*_path + L*_traj) in Eq. (9) is independent of K, and the paper never estimates C1, C2, or the achieved L* values. Moreover, Appendix C.6 states that Assumption 1(i) is 'assumed rather than explicitly imposed' and that no spectral normalization or other mechanism enforces a known Lipschitz constant. As a result, the bound cannot be checked numerically as presented, and the statement that small-K inference is 'a consequence of the training objective rather than an empirical heuristic' is stronger than the submitted result supports. The authors should either supply estimates or a mechanism for the constants, or explicitly downgrade the claim to a heuristic design principle with a proof sketch.
minor comments (6)
  1. [Section 3, Eq. (2), Table 3] The load-balancing loss notation is inconsistent: Eq. (2) uses lambda_I and lambda_L, while Eq. (12) and Table 3 refer to lambda_aux; please align these notations.
  2. [Appendix C.5] The sentence claiming that all sampling steps lie strictly in the interior of [0,T] contradicts t_K=T; please clarify that the endpoint is used only for initialization and state explicitly that the first reverse step is handled by a separate argument.
  3. [Figure 5] The ring axes in the scene-stratified plot are not labeled with units, and the caption does not explain the scaling of the inner versus outer rings; adding a brief axis explanation would improve readability.
  4. [Throughout] The name 'VoiceBank' is typeset with a stray space in several places (for example, in the Abstract and Section 4); this should be fixed.
  5. [Section 4, Table 1] Since all baseline numbers were reproduced by the authors, it would be helpful to report the number of evaluation runs and the variance of the metrics, so the reader can assess whether the reported differences are statistically meaningful.
  6. [Related Work] The paper states that this is 'the first joint MoE–SB framework for SE'; if this is intended as a novelty claim, please verify that no prior joint work exists and cite the closest alternatives.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: Theorem 1 has the standard training-loss-controlled sampling bound form; the boundary t=T proof gap is a correctness/missing-support issue, not a circular reduction, and the only self-citations are non-load-bearing related-work mentions.

full rationale

The paper's central theoretical claim is Theorem 1, which bounds the 2-Wasserstein distance between the K-step rollout and the continuous-time bridge marginal by C1*K^(-alpha) + C2*sqrt(L*_path + L*_traj). This is the standard structure of a posterior-sampling discretization bound: sampling error is controlled by training-time quantities that the optimization actually minimizes. The regularizers in Eqs. (7)-(8) are not defined in terms of the test-time W2 metric, and Eq. (9) is a derived inequality (modulo the stated regularity assumptions), so the conditional guarantee does not reduce to a fit. The fusion calibration loss in Eq. (11) anchors the epistemic and aleatoric scalars to reconstruction errors, but the reported PESQ, STOI, and ECE are computed on the held-out test set with external metrics, so that is not a circular prediction either. Two caveats keep the score from being zero: (a) the paper cites two works by a co-author, references [12] and [41], as related work, although neither is load-bearing for the new contributions; and (b) the proof of Lemma 1 in Appendix C.2 restricts the one-step fidelity bound to [epsilon0, T-epsilon0], while the K-step rollout of Eq. (5) first evaluates the denoiser at t=T, where L_traj is identically zero by Eq. (8) because beta_T=0 and sigma_T=0. The paper's statement in Appendix C.5 that 'all sampling steps lie strictly in the interior' is internally inconsistent with t_K=T. This is an omitted-support/proof gap in the boundary step, not a circular reduction: it does not make Eq. (9) equal to its input, but it weakens the claimed objective-level justification of K=8. Because the remaining contributions are externally benchmarked and parameter-free relative to the test metric, the circularity is minimal.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claims rest mainly on Assumption 1 in Appendix C, plus the choice of hyperparameters including sigma_max, gamma, K, and loss weights. No new physical entities are introduced. The theorem's practical force depends on the unverified Lipschitz assumption and on unmeasured constants and training losses.

free parameters (6)
  • sigma_max = 0.05
    Bridge perturbation scale chosen by the authors; controls the full-support marginal and the irreducible noise floor in L_traj, and appears in the non-vanishing term of Theorem 1.
  • gamma = 0.6
    Front-loading exponent used in the inference schedule t_k = T(k/K)^gamma and in the theoretical rate alpha = min(1, gamma). Chosen by hand, not estimated.
  • K = 8
    Inference sampling steps. The paper claims Theorem 1 justifies this choice, but C1, C2, and L* are not measured, so the justification is not numerically demonstrated.
  • Loss weights (lambda_SB, lambda_path, lambda_traj, lambda_aux, lambda_cal, lambda_I, lambda_L) = 1.0, 0.1, 0.05, 0.01, 0.05, 0.5, 0.5
    All loss weights are grid-searched on a 10% held-out validation split of VoiceBank+DEMAND, as stated in Section 3 and Table 3.
  • Ceilings M_max and phi_max = 5.0 and pi/4
    Mask and phase correction ceilings chosen to prevent over-suppression and phase instability; no sensitivity analysis is reported.
  • Gate blend alpha = not specified
    The two-level routing gate is defined as G = alpha*G_arch + (1-alpha)*G_token with alpha in [0,1], but no value is reported in the paper.
assumptions (4)
  • domain assumption Assumption 1(i): the trained denoiser hat_x_theta is Lx-Lipschitz in its first argument and Lt-Lipschitz in t.
    Stated in Appendix C and required for Theorem 1. Appendix C.6 admits this is 'assumed rather than explicitly imposed' and that no Lipschitz constant is enforced or estimated.
  • standard math Assumption 1(ii): beta_bar_t and sigma_t are C1 with beta_bar_t monotone and bounded second derivative.
    The cosine schedule is smooth on [0,T] and sigma_t has endpoint singularities that the proof handles by truncation; this is a standard regularity condition in SDE discretization analyses.
  • standard math Assumption 1(iii): second moments of x_t and hat_x_theta(xt,y,t) are uniformly bounded in t.
    Justified in Appendix C.5 using amplitude-normalized waveforms and the Lipschitz condition; this is a mild moment assumption.
  • domain assumption The forward construction of Eq. (4) with sigma_max = 0.05 yields full-support marginals required for score-based reverse sampling.
    Invoked in Section 3 to justify retaining the small noise term sigma_t*epsilon; the choice of sigma_max is a free parameter, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HybridSB-MoE: Dual-Domain Schr\"odinger Bridges with Scene-Adaptive Expert Routing for Speech Enhancement." pith.science (2026). https://pith.science/paper/M3GKZ7SA

@misc{pith2026260812715,
  author       = {Pith},
  title        = {Pith review of: HybridSB-MoE: Dual-Domain Schr\"odinger Bridges with Scene-Adaptive Expert Routing for Speech Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3GKZ7SA}},
  note         = {Machine review of arXiv:2608.12715}
}
read the original abstract

Generative speech enhancement faces three gaps: spectral models capture harmonic structure but often disrupt phase, waveform models preserve phase but miss harmonics, and Schr\"odinger Bridges (SB) shorten transport from noise to clean speech but leave inference cost only loosely tied to training. We propose HybridSB-MoE, a dual-domain framework that fills these gaps through three contributions unified by a single asymmetric design principle. (i) Asymmetric uncertainty fusion: The spectral path captures epistemic uncertainty via expert disagreement, while the waveform bridge models aleatoric variance through stochastic dynamics. We fuse them asymmetrically, allowing the mixing weight to adapt to distinct error regimes rather than average predictions. (ii) Heterogeneous MoE with top-k=2 routing across five distinct architectural archetypes, where architectural diversity makes the epistemic signal indicate which inductive bias fails rather than small perturbations among similar experts. (iii) Discretization bound (Theorem 1): path-consistency and trajectory regularizers together bound the K-step bridge sampling error in 2-Wasserstein distance at rate K-alpha, making small-K inference an objective-level guarantee rather than an empirical claim. On VoiceBank+DEMAND, HybridSB-MoE outperforms diffusion- and SB-based baselines at their step budgets while remaining competitive with consistency-distilled few-step methods.

Figures

Figures reproduced from arXiv: 2608.12715 by the authors.

Figure 1
Figure 1. The persistent dual-domain di￾chotomy in speech enhancement. Wave￾form processing preserves temporal fine structure and phase coherence but under￾performs on harmonically structured in￾terference; spectral processing captures harmonic structure and stationary noise patterns but fragments phase across STFT frames. The two domains succeed and fail in opposite regimes, and existing genera￾tive methods inherit one half … view at source ↗
Figure 2
Figure 2. Overview of HybridSB-MoE. The spectral pathway (top) routes log-magnitude features [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Ablation studies and efficiency analysis. (a) Each architectural component contributes [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (reproduced here for completeness) illustrates the shared backbone and the architecturally distinct instantiations. All experts pass features through LayerNorm → variable internal modules → bottleneck projection, ensuring interface compatibility for routing while permi…
Figure 5
Figure 5. Figure 5: Scene-stratified performance across all 14 noise types in VoiceBank+DEMAND. Inner [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Spectrogram comparison before and after enhancement. HybridSB-MoE attenuates [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 35 canonical work pages

  1. [1]

    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

  2. [2]

    Real time speech enhancement in the waveform domain,

    A. Defossez, G. Synnaeve, and Y . Adi, “Real time speech enhancement in the waveform domain,” arXiv preprint arXiv:2006.12847, 2020

    arXiv 2006

  3. [3]

    A regression approach to speech enhancement based on deep neural networks,

    Y . Xu, J. Du, L.-R. Dai, and C.-H. Lee, “A regression approach to speech enhancement based on deep neural networks,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 1, pp. 7–19, 2015

    work page 2015

  4. [4]

    Segan: Speech enhancement generative adversarial network,

    S. Pascual, A. Bonafonte, and J. Serra, “Segan: Speech enhancement generative adversarial network,”arXiv preprint arXiv:1703.09452, 2017

    arXiv 2017

  5. [5]

    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 Trans. Audio, Speech and Lang. Proc., vol. 27, no. 8, p. 1256–1266, Aug. 2019. [Online]. Available: https://doi.org/10.1109/TASLP.2019.2915167

    arXiv 2019

  6. [6]

    Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation,

    Y . Luo, Z. Chen, and T. Yoshioka, “Dual-path rnn: efficient long sequence modeling for time-domain single-channel speech separation,” inICASSP, 2020, pp. 46–50

    work page 2020

  7. [7]

    Fullsubnet+: Channel attention fullsubnet with complex spectrograms for speech enhancement,

    J. Chen, Z. Wang, D. Tuo, Z. Wu, S. Kang, and H. Meng, “Fullsubnet+: Channel attention fullsubnet with complex spectrograms for speech enhancement,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7857–7861

    work page 2022

  8. [8]

    Tf-gridnet: Integrating full- and sub-band modeling for speech separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y . Lee, B.-Y . Kim, and S. Watanabe, “Tf-gridnet: Integrating full- and sub-band modeling for speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 3221–3236, 2023

    work page 2023

Show all 55 references
  1. [9]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inProceedings of the 34th International Conference on Neural Information Processing Systems, ser. NIPS ’20. Red Hook, NY , USA: Curran Associates Inc., 2020

    2020

  2. [10]

    Speech enhancement with score-based generative models in the complex stft domain,

    S. Welker, J. Richter, and T. Gerkmann, “Speech enhancement with score-based generative models in the complex stft domain,”arXiv preprint arXiv:2203.17004, 2022

    2022 arXiv

  3. [11]

    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 Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2351–2364, 2023

    2023

  4. [12]

    Sparse mixture of local experts for efficient speech enhancement,

    A. Sivaraman and M. Kim, “Sparse mixture of local experts for efficient speech enhancement,” arXiv preprint arXiv:2005.08128, 2020

    2005 arXiv

  5. [13]

    Speech enhancement with mixture of deep experts with clean clustering pre-training,

    S. E. Chazan, J. Goldberger, and S. Gannot, “Speech enhancement with mixture of deep experts with clean clustering pre-training,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 716–720

    2021

  6. [14]

    Suppression of acoustic noise in speech using spectral subtraction,

    S. Boll, “Suppression of acoustic noise in speech using spectral subtraction,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 27, no. 2, pp. 113–120, 1979

    1979

  7. [15]

    Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator,

    Y . Ephraim and D. Malah, “Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator,”IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 32, no. 6, pp. 1109–1121, 1984

    1984

  8. [16]

    Sdr – half-baked or well done?

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

    2019

  9. [17]

    An investigation of incorporating mamba for speech enhancement,

    R. Chao, W.-H. Cheng, M. La Quatra, S. M. Siniscalchi, C.-H. H. Yang, S.-W. Fu, and Y . Tsao, “An investigation of incorporating mamba for speech enhancement,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 302–308. 10

    2024

  10. [18]

    Mamba-seunet: Mamba unet for monaural speech enhancement,

    J. Wang, Z. Lin, T. Wang, M. Ge, L. Wang, and J. Dang, “Mamba-seunet: Mamba unet for monaural speech enhancement,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  11. [19]

    Diffusion-based speech enhancement with joint generative and predictive decoders,

    H. Shi, K. Shimada, M. Hirano, T. Shibuya, Y . Koyama, Z. Zhong, S. Takahashi, T. Kawahara, and Y . Mitsufuji, “Diffusion-based speech enhancement with joint generative and predictive decoders,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal...

    2024

  12. [20]

    Speech enhancement algorithms: A systematic literature review,

    S. T. Yousif and B. M. Mahmmod, “Speech enhancement algorithms: A systematic literature review,”Algorithms, vol. 18, no. 5, p. 272, 2025

    2025

  13. [21]

    Likelihood training of schrödinger bridge using forward–backward sdes theory,

    T. Chen, G.-H. Liu, and E. A. Theodorou, “Likelihood training of schrödinger bridge using forward–backward sdes theory,” inICLR, 2022

    2022

  14. [22]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inInternational Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id= PxTIG12RRHS

    2021

  15. [23]

    Denoising diffusion implicit models,

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

    2010 arXiv

  16. [24]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inInternational Conference on Learning Representations, 2023

    2023

  17. [25]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inInternational Conference on Learning Representations, 2023

    2023

  18. [26]

    Diffusion Schrödinger bridge with applications to score-based generative modeling,

    V . De Bortoli, J. Thornton, J. Heng, and A. Doucet, “Diffusion Schrödinger bridge with applications to score-based generative modeling,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 17 695–17 709

    2021

  19. [27]

    Diffusion schrödinger bridge matching,

    Y . Shi, V . De Bortoli, A. Campbell, and A. Doucet, “Diffusion schrödinger bridge matching,” inAdvances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 62 183–62 2...

    2023

  20. [28]

    Schrödinger bridge for generative speech enhancement,

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

    2024 arXiv

  21. [29]

    Diffusion-based speech enhancement with Schrödinger bridge and symmetric noise schedule,

    S. Wang, S. Liu, A. Harper, P. Kendrick, M. Salzmann, and M. Cernak, “Diffusion-based speech enhancement with Schrödinger bridge and symmetric noise schedule,”arXiv preprint arXiv:2409.05116, 2024

    2024 arXiv

  22. [30]

    Simplified diffusion schrödinger bridge,

    Z. Tang, T. Hang, S. Gu, D. Chen, and B. Guo, “Simplified diffusion schrödinger bridge,”arXiv preprint arXiv:2403.14623, 2024

    2024 arXiv

  23. [31]

    Schrödinger bridge mamba for one-step speech enhancement,

    J. Yang, S. Wang, C. Wu, L. Guo, and F. Fan, “Schrödinger bridge mamba for one-step speech enhancement,”arXiv preprint arXiv:2510.16834, 2025

    2025 arXiv

  24. [32]

    Sb-senet: Diffusion model based on schrödinger bridge for speech enhancement,

    H. Zhang, G. Li, P. Wu, Y . Gao, and H. Zhang, “Sb-senet: Diffusion model based on schrödinger bridge for speech enhancement,”Applied Acoustics, vol. 236, p. 110742, 2025

    2025

  25. [33]

    Robust one-step speech enhancement via consistency distillation,

    L. Xu, L. F. Yan, and W. B. Kleijn, “Robust one-step speech enhancement via consistency distillation,”arXiv preprint arXiv:2507.05688, 2025

    2025 arXiv

  26. [34]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,” inInternational Conference on Machine Learning, 2023

    2023

  27. [35]

    Few-step adversarial schrödinger bridge for generative speech enhancement,

    S. Han, S. Lee, J. Lee, and K. Lee, “Few-step adversarial schrödinger bridge for generative speech enhancement,”arXiv preprint arXiv:2506.01460, 2025. 11

    2025 arXiv

  28. [36]

    Schrödinger bridge consistency trajectory models for speech enhancement,

    S. Nishigori, K. Saito, N. Murata, M. Hirano, S. Takahashi, and Y . Mitsufuji, “Schrödinger bridge consistency trajectory models for speech enhancement,”arXiv preprint arXiv:2507.11925, 2025

    2025 arXiv

  29. [37]

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions,

    S. Chen, S. Chewi, J. Li, Y . Li, A. Salim, and A. R. Zhang, “Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions,” inInternational Conference on Learning Representations, 2023

    2023

  30. [38]

    Outra- geously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outra- geously large neural networks: The sparsely-gated mixture-of-experts layer,”arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  31. [39]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,”Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022. [Online]. Available: http://jmlr.org/papers/v23/21-0998.html

    2022

  32. [40]

    Gshard: Scaling giant models with conditional computation and automatic sharding,

    D. Lepikhin, H. Lee, Y . Xu, D. Chen, O. Firat, Y . Huang, M. Krikun, N. Shazeer, and Z. Chen, “Gshard: Scaling giant models with conditional computation and automatic sharding,”arXiv preprint arXiv:2006.16668, 2020

    2006 arXiv

  33. [41]

    Zero-shot personalized speech enhancement through speaker- informed model selection,

    A. Sivaraman and M. Kim, “Zero-shot personalized speech enhancement through speaker- informed model selection,” in2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2021, pp. 171–175

    2021

  34. [42]

    Adaptive slimming for scalable and efficient speech enhancement,

    R. Miccini, M. Kim, C. Laroche, L. Pezzarossa, and P. Smaragdis, “Adaptive slimming for scalable and efficient speech enhancement,”arXiv preprint arXiv:2507.04879, 2025

    2025

  35. [43]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  36. [44]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the 34th International Conference on Machine Learning - Volume 70, ser. ICML’17. JMLR.org, 2017, p. 1321–1330

    2017

  37. [45]

    A unified approach to short-time fourier analysis and synthesis,

    J. B. Allen and L. R. Rabiner, “A unified approach to short-time fourier analysis and synthesis,” Proceedings of the IEEE, vol. 65, no. 11, pp. 1558–1564, 1977

    1977

  38. [46]

    The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The diverse environments multi-channel acoustic noise database (demand): A database of multichannel environmental noise recordings,” Proceedings of Meetings on Acoustics, vol. 19, no. 1, p. 035081, 05 2013. [Online]. Available: https://doi...

    2013 doi

  39. [47]

    Group normalization,

    Y . Wu and K. He, “Group normalization,” inProceedings of the European Conference on Computer Vision, 2018, pp. 3–19

    2018

  40. [48]

    Computational optimal transport,

    G. Peyre and M. Cuturi, “Computational optimal transport,”Foundations and Trends in Machine Learning, vol. 11, no. 5-6, pp. 355–607, 2019

    2019

  41. [49]

    Improved denoising diffusion probabilistic models,

    A. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” 2021. [Online]. Available: https://arxiv.org/abs/2102.09672

    2021 arXiv

  42. [50]

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

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Investigating rnn-based speech enhancement methods for noise-robust text-to-speech,” in9th ISCA Workshop on Speech Synthesis Workshop (SSW 9), 2016, pp. 146–152

    2016

  43. [51]

    Perceptual evaluation 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 evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and codecs,” in2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No...

    2001

  44. [52]

    An algorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An algorithm for intelligibility prediction of time–frequency weighted noisy speech,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 7, pp. 2125–2136, 2011. 12

    2011

  45. [53]

    Evaluation of objective quality measures for speech enhancement,

    Y . Hu and P. C. Loizou, “Evaluation of objective quality measures for speech enhancement,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 16, no. 1, pp. 229–238, 2008

    2008

  46. [54]

    Deep learning and the information bottleneck principle,

    N. Tishby and N. Zaslavsky, “Deep learning and the information bottleneck principle,” in2015 IEEE Information Theory Workshop (ITW), 2015, pp. 1–5. 13 A Heterogeneous Expert Architectures This appendix details the heterogeneous MoE layer described in Section 3 of the main pape...

    2015

  47. [55]

    Y-axis ranges are zoomed for visibility of cross-scene variation

    achieves a mean PESQ of 3.85 across these conditions; our method exceeds this in every scene, with PESQ ranging from 3.84 to 3.92 (mean 3.88, standard deviation<0.03 across scenes). Y-axis ranges are zoomed for visibility of cross-scene variation. observationy). Both moments a...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.