Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

DeepFilterGAN: A Full-band Real-time Speech Enhancement System with GAN-based Stochastic Regeneration

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 DeepFilterGAN claims that a full-band, real-time speech enhancer can improve perceived quality by regenerating speech from a two-channel input of noisy and intermediate-enhanced spectrograms.

desk verdict A competent engineering paper with a clean noisy-conditioning ablation, but the headline MOS gain is measured against a weakened retrained first stage, so the broader claim about improving a strong predictive enhancer is not yet supported. read the letter →

arxiv 2505.23515 v1 pith:33KZTKGK submitted 2025-05-29 eess.AS cs.LGeess.SP

Sanberk Serbest , Tijana Stojkovic , Milos Cernak , Andrew Harper This is my paper
classification eess.AScs.LGeess.SP
keywords speechenhancementstochasticregenerationgenerativeadversarialnetworkreal-timeprocessingfull-bandaudiodeepfilteringnoisyconditioningURGENTchallenge
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

DeepFilterGAN is a full-band speech enhancement system that pairs a predictive first stage with a generative adversarial second stage. The paper's central claim is that this second stage, trained with the first stage frozen, improves perceived speech quality by regenerating the clean speech distribution from a two-channel input of the noisy signal and the first stage's output. On the 2024 URGENT non-blind test set, the system raises NISQA-MOS from 2.66 (first stage alone) to 3.12, with 3.58M parameters and 40 ms algorithmic latency. An ablation that removes the noisy input drops to 2.86, which the authors present as evidence that noisy conditioning is what enables the gain. This matters because generative restoration has usually required large models or many inference steps; here it fits inside a streaming, full-band system.

What carries the argument

Stochastic regeneration with two-channel conditioning. The first stage produces an intermediate enhanced spectrogram $Z(k,f)$; the generator receives $[Y(k,f), Z(k,f)]$ as a two-channel input and is trained to sample the clean speech distribution $P(x|y,z)$ under a multi-scale adversarial loss. The generator is a lightweight modification of Online SpatialNet (1.14M parameters) and the discriminator follows MelGAN's three-scale design; the final objective is $\min_G \mathbb{E}_{y,z}[\sum_l -D_l(\hat{x})] + \beta\|x-\hat{x}\|_1$, where $\hat{x} = \operatorname{istft}(G(y,z))$.

What would settle it

Run the two variants on test utterances where the first stage removes entire phonemes; if the no-concat variant still recovers those phonemes, or if replacing the noisy channel with a fixed constant keeps NISQA-MOS near 3.12, the paper's noisy-conditioning mechanism is not what carries the gain.

Watch

Extended reading notes

Core claim

The discovery is that stochastic regeneration with a GAN can repair over-suppression in a real-time predictive enhancer. Specifically, the generator receives the concatenated spectrograms $[Y(k,f), Z(k,f)]$ and is trained against three MelGAN-style discriminators with a hinge loss plus a time-domain L1 term, learning the conditional distribution $P(x|y,z)$. On the challenge's non-blind test set, the final system scores 3.12 NISQA-MOS versus 2.66 for the retrained first stage, holds PESQ, ESTOI, SDR, and LSD roughly constant, and obtains the best overall ranking among compared models. A spectrogram example shows a speech segment removed by the first stage partially reappearing only when the noisy input is included.

Load-bearing premise

The load-bearing premise is that a 1.14M-parameter generator can learn the clean-speech distribution from a two-channel input of noisy and first-stage-enhanced spectrograms, so that the measured NISQA-MOS gain is genuine restoration rather than a generic reshaping of the output.

Editorial extensions

If this is right

  • The full system improves NISQA-MOS from 2.66 to 3.12 over the retrained first stage on the 2024 URGENT non-blind test set, while other objective metrics stay roughly level.
  • Concatenating the noisy input with the first-stage output is load-bearing for the improvement: removing it lowers NISQA-MOS to 2.86 and severely reduces recovery of over-suppressed speech content.
  • A generative second stage can fit inside a real-time system: 3.58M training parameters (3.45M at inference) and 40 ms algorithmic latency.
  • Despite lower NISQA-MOS than a larger diffusion-based comparison system, DeepFilterGAN obtains the best overall ranking across nonintrusive, intrusive, downstream-task-independent, and downstream-task-dependent metric categories.
  • The system preserves intelligibility-side metrics (phoneme similarity and word accuracy rate) better than the larger diffusion-based comparison.

Reading between the lines

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

  • The noisy conditioning channel may act as more than noise context: it could anchor the generator to the input signal, and a scrambling experiment could separate those two roles.
  • Jointly training both stages, which the paper lists as future work, is a natural next step because a frozen first stage may limit how well the generator can learn $P(x|y,z)$; end-to-end training could widen the NISQA-MOS gap.
  • Because the second stage has only 1.14M parameters, the same noisy-concat conditioning could be tested on even smaller on-device models.
  • The time-domain L1 term in the generator objective probably explains why intrusive metrics stay stable; varying $\beta$ would directly probe the trade-off between adversarial sharpness and signal conservation.
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 DeepFilterGAN, a two-stage causal speech enhancement system. The first stage is a retrained DeepFilterNet2; the second stage is a GAN whose generator is based on OnlineSpatialNet and whose discriminator is based on MelGAN. The generator receives the concatenation of the noisy and intermediate-enhanced STFTs, which the authors call noisy conditioning. The system has 3.58M parameters and 40 ms algorithmic latency. Experiments on the 2024 URGENT non-blind test set report NISQA-MOS of 3.12 for the proposed system versus 2.66 for the retrained first stage and 2.86 for an ablation without noisy conditioning. The paper also compares with published DeepFilterNet2 (3.28) and UNIVERSE++ (3.44) and reports a favorable overall ranking.

Significance. If the central result holds, the paper demonstrates that a compact GAN second stage can improve a predictive enhancer's perceived quality without requiring multiple diffusion inference steps, which is useful for real-time and embedded speech enhancement. The evaluation uses an external challenge test set with multiple objective metrics, and the ablation isolates the contribution of noisy conditioning. Strengths include the low parameter count, the explicit algorithmic-latency statement, and the use of a public challenge protocol. However, the main empirical claim is weakened by the unexplained gap between the retrained first stage and the published DeepFilterNet2 checkpoint, and by missing statistical and reproducibility details.

major comments (4)
  1. [Table 1, Section 4] The baseline used to demonstrate the central improvement, 'First Stage only (retrained [11])', obtains NISQA-MOS 2.66, which is 0.62 below the publicly released DeepFilterNet2 checkpoint (3.28) evaluated on the same test set. Since the first stage is explicitly adopted from DeepFilterNet2, the paper should explain why retraining on the 2025 URGENT training set yields such a large drop; the training description in Section 3.2 reports only '45 epochs' and omits the loss weights and optimizer details. Because the second stage is trained on top of the retrained first stage, the reported gain (2.66 to 3.12) may be partly compensating for an underfit first stage. The published DeepFilterNet2 checkpoint is available, so the authors can directly test this by training the identical second stage on the published first stage and reporting the resulting NISQA-MOS; alternatively, they should show that the retrained first stage is not anomalously weak.
  2. [Sections 3.1 and 3.2, Eq. (3)] The second-stage generator is described only as 'adopted from OnlineSpatialNet' with two input channels, four blocks, and hidden size 16. The paper does not state whether the generator outputs a complex mask, a complex spectrogram, or a magnitude estimate; it also omits the value of beta in Eq. (3), the learning rates for generator and discriminator, the gradient-clipping threshold, and the exact weight-normalization and update schedules beyond 'every second iteration'. Without these details, the experiments are not reproducible and the contribution of the specific generator design to the NISQA gain cannot be isolated. Please provide an architecture summary (layer types, kernel sizes, skip connections, Mamba configuration) and a complete hyperparameter table.
  3. [Table 1 and Section 4] All metrics in Table 1 are single-point estimates with no confidence intervals or significance tests. The central claim relies on a NISQA-MOS increase of 0.46 over the first stage and an ablation difference of 0.26 for noisy concatenation, but NISQA is a non-intrusive predictor with nontrivial prediction variance, and several objective metrics in the same table move in the opposite direction (e.g., PESQ 2.07 to 2.03, word accuracy 75.51 to 75.05). At minimum, the authors should report bootstrap confidence intervals or per-utterance distributions for the principal metrics (especially NISQA-MOS, PESQ, and word accuracy) to establish that the reported improvements and non-degradations are not within the noise.
  4. [Section 3.2 and title] The title and abstract describe the system as real-time, but the paper reports only the 40 ms algorithmic latency from the two-frame lookahead; no real-time factor (RTF) or measured inference latency on any hardware is provided. For a streaming/embedded system, the RTF relative to real time on the target platform is a standard requirement and is load-bearing for the 'real-time' claim. Please report the RTF or wall-clock latency for at least one representative device (e.g., a notebook CPU or an embedded ARM core).
minor comments (5)
  1. [Section 2.4] The text states that each MelGAN discriminator 'focuses on a different frequency range', but the original MelGAN multi-scale discriminators operate on different time scales (via sub-sampling of the waveform), not frequency ranges. Please correct the description.
  2. [Section 3.2] The sentence 'We use the same learning rate and weight decay schedule as in [11]' is ambiguous: it appears after the description of second-stage training, but it is unclear whether it applies to the first-stage training (which is described only by '45 epochs') or only to the GAN training. Please state explicitly which learning-rate and weight-decay schedules are used for each training step.
  3. [Table 1] The row label 'First Stage only (retrained [11])' is confusing because the published DeepFilterNet2 row is also present. Rename it to something like 'First stage only (retrained by authors)' to make the distinction explicit.
  4. [Section 4] The overall ranking in Eq. (4) is computed over only five systems, so the difference between ranking scores (e.g., 2.25 vs 2.31) is coarse; the text should note that the ranking interpretation is limited by the small number of compared systems.
  5. [Section 4, Figure 3] The green-boxed region in Figure 3 is described only by color; if the manuscript is printed in grayscale, the comparison is hard to follow. Add a text label or a distinct line style to mark the recovered segment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central NISQA-MOS comparison is an external, empirical evaluation and is not constructed from fitted parameters or self-citations.

full rationale

The paper's central claim is an empirical performance improvement of the proposed two-stage system over its first stage, measured on the 2024 URGENT non-blind test set with external objective metrics such as NISQA-MOS, PESQ, ESTOI, SDR, and LSD. The second-stage generator is trained with the objective in Equation 3, which combines an adversarial hinge loss with a time-domain L1 term; no parameter in this objective or in the architecture is defined in terms of the reported evaluation metrics, and no fitted quantity is later relabeled as a prediction. The ablation study compares system variants (with and without noisy concatenation) on the same external test set, which is a standard empirical comparison rather than a circular derivation. The paper contains no load-bearing self-citations: references [11], [12], and [13] are external prior works whose architectures and pretrained checkpoints are reused, but the central result does not reduce to an assertion of those works. The only circularity-adjacent concern is that the retrained first-stage baseline (2.66 NISQA-MOS) is considerably weaker than the published DeepFilterNet2 checkpoint (3.28), which could weaken the force of the improvement claim; however, this is a benchmarking or experimental-design issue, not a circular one, because the measured gain is still an external empirical outcome rather than an identity. Overall, no specific reduction by construction, no fitted input renamed as a prediction, and no self-citation chain supporting the main result were found.

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

The central claim depends on hand-chosen hyperparameters (most importantly the unreported beta in Equation 3) and on several architectural and training assumptions inherited from prior work. No new physical entities, forces, or constants are introduced. The empirical comparison rests on the validity of benchmark metrics and on the fairness of comparing a model trained on URGENT data against external baselines trained on other datasets.

free parameters (6)
  • beta (beta in Eq. 3) = not reported
    Weighting between adversarial loss and time-domain L1 loss in the generator objective. Chosen by hand or tuned, but its value is not given, and it directly controls the fidelity versus distribution-matching tradeoff.
  • generator hidden size = 16
    Hidden size for the Online SpatialNet-based generator; selected to obtain a 1.14M parameter second stage. It determines the capacity to learn the speech distribution.
  • generator number of blocks = 4
    Number of Mamba blocks in the generator; a hand-chosen architectural parameter that controls receptive field and model capacity.
  • GAN training epochs = 200
    Number of epochs for second-stage training. The stopping point is hand-chosen and may affect the reported metrics.
  • discriminator update interval = every 2 iterations
    The discriminator is updated every second iteration to stabilize adversarial training; a hand-set training heuristic.
  • lookahead frames = 2 frames (40 ms total latency)
    Causal lookahead in the STFT domain; hand-chosen latency-quality tradeoff that defines the real-time claim.
assumptions (4)
  • domain assumption The two-channel concatenation of noisy and intermediate enhanced spectrograms provides sufficient conditioning to learn P(x|y,z).
    Section 3.1 and Figure 2: the generator is conditioned on Y and Z, with the claim that this carries both noise information and over-suppressed speech content. No proof or analysis shows this is sufficient.
  • domain assumption Training the first stage frozen and training only the second stage on the same dataset is a stable and effective realization of stochastic regeneration.
    Section 3.2: the first stage is pre-trained and frozen. The paper lists joint training as future work, acknowledging this setup may be suboptimal.
  • domain assumption GAN training with three MelGAN discriminators on time-domain signals improves perceived quality without introducing damaging artifacts.
    Section 3.2: the adversarial objective is used to push the generated distribution toward clean speech. The paper relies on metrics like NISQA-MOS and PESQ to validate this, but no artifact analysis or listening test is provided.
  • domain assumption The objective metrics used (NISQA-MOS, PESQ, ESTOI, SDR, LSD, phoneme similarity, word accuracy) are valid proxies for speech quality and intelligibility.
    Section 4: these metrics are adopted from the URGENT Challenge. The paper does not discuss their limitations or provide human evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepFilterGAN: A Full-band Real-time Speech Enhancement System with GAN-based Stochastic Regeneration." pith.science (2026). https://pith.science/paper/33KZTKGK

@misc{pith2026250523515,
  author       = {Pith},
  title        = {Pith review of: DeepFilterGAN: A Full-band Real-time Speech Enhancement System with GAN-based Stochastic Regeneration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/33KZTKGK}},
  note         = {Machine review of arXiv:2505.23515}
}
read the original abstract

In this work, we propose a full-band real-time speech enhancement system with GAN-based stochastic regeneration. Predictive models focus on estimating the mean of the target distribution, whereas generative models aim to learn the full distribution. This behavior of predictive models may lead to over-suppression, i.e. the removal of speech content. In the literature, it was shown that combining a predictive model with a generative one within the stochastic regeneration framework can reduce the distortion in the output. We use this framework to obtain a real-time speech enhancement system. With 3.58M parameters and a low latency, our system is designed for real-time streaming with a lightweight architecture. Experiments show that our system improves over the first stage in terms of NISQA-MOS metric. Finally, through an ablation study, we show the importance of noisy conditioning in our system. We participated in 2025 Urgent Challenge with our model and later made further improvements.

Figures

Figures reproduced from arXiv: 2505.23515 by the authors.

Figure 1
Figure 1. Stochastic regeneration. y is the noisy speech, z is the intermediate enhanced speech and xˆ is the final enhanced speech, i.e. the clean speech estimate of the overall system. Different approaches have been investigated in combining the predictive and generative approaches. In [5], a stochastic regeneration system is proposed by using a diffusion model for denoising and dereverberation tasks. In [8], an extended Po… view at source ↗
Figure 2
Figure 2. Our proposed model. Y (k, f) is the STFT of noisy speech, Z(k, f) is the STFT of the intermediate enhanced speech and Xˆ(k, f) is the STFT of the final enhanced speech, i.e. the clean speech estimate of the overall system. 3.1. Networks The first stage predictive model is DeepFilterNet2 from [11]. It comprises of one encoder and two decoders containing mainly convolutional layers, grouped linear layers and gated rec… view at source ↗
Figure 3
Figure 3. The recovery performance of our proposed system. The area in the green box is removed in the first stage output. Our system with noisy concatenation recovers some portion of this segment while the model without noisy concatenation can’t recover it. UNIVERSE++ [6] trained on Voicebank-DEMAND [17] and DeepFilternet2 [11] trained on DNS Challenge [18] dataset. We also provide an ablation study to discuss the improvemen… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Leveraging Discriminative Latent Representations for Conditioning GAN-Based Speech Enhancement

    eess.AS 2025-08 conditional novelty 4.0 of 10

    DisCoGAN uses latent features from a frozen discriminative speech-enhancement encoder as conditioning for a GAN generator, improving low-SNR enhancement over existing GAN pipelines.

Reference graph

Works this paper leans on

40 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [11]

    DeepFilterNet2: Towards real-time speech enhancement on em- bedded devices for full-band audio,

    H. Schr ¨oter, A. N. Escalante-B., T. Rosenkranz, and A. Maier, “DeepFilterNet2: Towards real-time speech enhancement on em- bedded devices for full-band audio,” in 17th International Work- shop on Acoustic Signal Enhancement (IWAENC 2022), 2022

    work page 2022

  2. [1]

    Introduction In the deep learning based speech enhancement (SE) litera- ture, there are many approaches including RNN-based meth- ods [1], transformer-based methods [2], Generative Adversarial Network (GAN)-based methods [3] and diffusion-based models [4, 5]. Many works directly focus on denoising or dereverbera- tion tasks [4] while there is also an atte...

    work page

  3. [2]

    Background and Related Works A conventional single channel speech enhancement system tries to recover the clean speech signal x(t) by observing the noisy speech signal y(t) =h(t) ∗ x(t) +n(t) where h(t) is the im- pulse response of the channel andn(t) is an additive noise. This relation can also be represented in short-time Fourier Transform (STFT) domain...

    arXiv 2025

  4. [3]

    Our system aims to enhance the performance of the first-stage model, DeepFilterNet2 [11], by incorporating a GAN-based second stage

    Proposed Method We propose a full-band real-time stochastic regeneration sys- tem combining a predictive model with a generative adversarial network (GAN). Our system aims to enhance the performance of the first-stage model, DeepFilterNet2 [11], by incorporating a GAN-based second stage. The second stage is designed to recover over-suppressed speech segme...

    work page

  5. [4]

    Mel-spectrogram L1 loss to ensure consistency with the clean speech distribution. We then concatenate the noisy speech spectrogramY(k, f ) and the intermediate enhanced spectrogram Z(k, f ) obtained from the first stage to feed them into the second stage of our system. During second-stage training, the generator and dis- criminator compete using a hinge l...

    work page

  6. [5]

    It provides noise information that led to the particular inter- mediate enhanced speech, which inherently contains a rela- tionship between the noise information and the behavior of the first stage model

    work page

  7. [6]

    As a result, the generator learns the clean speech distribu- tion conditioned on the noisy speech and the intermediate en- hanced speech, i.e

    It provides the speech content that is possibly removed dur- ing the first stage process, which can be useful during the generation. As a result, the generator learns the clean speech distribu- tion conditioned on the noisy speech and the intermediate en- hanced speech, i.e. P(x|y, z). Our overall system that processes 48kHz speech samples in STFT domain ...

    work page 2024

  8. [7]

    Spectral loss and multi-resolution spectrogram loss to pre- serve the frequency structure,

    work page

Show all 40 references
  1. [8]

    Local SNR loss to maintain speech intelligibility,

  2. [9]

    In addition to the output speech quality, another aspect of speech enhancement models is the deployment requirements

    uses a GAN in a similar setting. In addition to the output speech quality, another aspect of speech enhancement models is the deployment requirements. For some speech enhancement applications such as mobile telecommunication, the SE system must be able to process streaming dat...

  3. [10]

    SI-SDR loss to enhance signal quality,

  4. [12]

    Experiments and Results We trained our proposed system with the training dataset of the 2025 Urgent Challenge [16]. This dataset is obtained from a collection of different individual datasets and contains speech samples with various sampling rates, namely 8k, 16k, 22.05k, 24k,...

    2025

  5. [13]

    The discriminator learns to distinguish real clean speech from enhanced speech, while the generator refines its output to resemble clean speech more closely

    while first stage model is frozen. The discriminator learns to distinguish real clean speech from enhanced speech, while the generator refines its output to resemble clean speech more closely. The optimization objectives are defined as follows: min Dl Ex [min(0, 1 − Dl(x))] +E...

  6. [14]

    The generative capabilities of the second stage and conditioning on the noisy speech along with the output of the first stage recover the over-suppressed speech content

    Conclusion In this work, we propose a full-band real-time stochastic regen- eration system combining a predictive model with a GAN called DeepFilterGAN. The generative capabilities of the second stage and conditioning on the noisy speech along with the output of the first stag...

  7. [15]

    High fidelity speech enhancement with band-split rnn,

    J. Yu, Y . Luo, H. Chen, R. Gu, and C. Weng, “High fidelity speech enhancement with band-split rnn,” 2023. [Online]. Available: https://arxiv.org/abs/2212.00406

    2023 arXiv

  8. [16]

    D2former: A fully complex dual-path dual- decoder conformer network using joint complex masking and complex spectral mapping for monaural speech enhancement,

    S. Zhao and B. Ma, “D2former: A fully complex dual-path dual- decoder conformer network using joint complex masking and complex spectral mapping for monaural speech enhancement,”

  9. [17]

    CMGAN: Conformer-based Metric GAN for Speech Enhancement,

    R. Cao, S. Abdulatif, and B. Yang, “CMGAN: Conformer-based Metric GAN for Speech Enhancement,” in Proc. Interspeech 2022, 2022, pp. 936–940

    2022

  10. [18]

    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

  11. [19]

    Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,

    J.-M. Lemercier, J. Richter, S. Welker, and T. Gerkmann, “Storm: A diffusion-based stochastic regeneration model for speech en- hancement and dereverberation,”IEEE/ACM Transactions on Au- dio, Speech, and Language Processing , vol. 31, pp. 2724–2737, 2023

    2023

  12. [20]

    Universal score-based speech enhancement with high content preservation,

    R. Scheibler, Y . Fujita, Y . Shirahata, and T. Komatsu, “Universal score-based speech enhancement with high content preservation,”

  13. [21]

    Performance measure- ment in blind audio source separation,

    E. Vincent, R. Gribonval, and C. Fevotte, “Performance measure- ment in blind audio source separation,”IEEE Transactions on Au- dio, Speech, and Language Processing, vol. 14, no. 4, pp. 1462– 1469, 2006

    2006

  14. [22]

    Schr ¨odinger bridge for generative speech enhancement,

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

    2024

  15. [23]

    Pfgm++ combined with stochastic regener- ation for speech enhancement,

    X. Cao and S. Zhao, “Pfgm++ combined with stochastic regener- ation for speech enhancement,” in 2024 9th International Confer- ence on Signal and Image Processing (ICSIP), 2024, pp. 267–271

    2024

  16. [24]

    Spectral oversubtraction? an approach for speech enhancement after robot ego speech filtering in semi-real-time,

    Y . Li, K. V . Hindriks, and F. A. Kunneman, “Spectral oversubtraction? an approach for speech enhancement after robot ego speech filtering in semi-real-time,” 2024. [Online]. Available: https://arxiv.org/abs/2409.06274

    2024 arXiv

  17. [25]

    Combined generative and predictive modeling for speech super-resolution,

    H. Wang, E. W. Healy, and D. Wang, “Combined generative and predictive modeling for speech super-resolution,” 2024. [Online]. Available: https://arxiv.org/abs/2401.14269

    2024 arXiv

  18. [27]

    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,” 2024. [Online]. Available: https://arxiv.org/abs/2403.07675

    2024 arXiv

  19. [28]

    Melgan: Generative adversarial networks for conditional waveform synthesis,

    K. Kumar, R. Kumar, T. de Boissiere, L. Gestin, W. Z. Teoh, J. Sotelo, A. de Brebisson, Y . Bengio, and A. Courville, “Melgan: Generative adversarial networks for conditional waveform synthesis,” 2019. [Online]. Available: https://arxiv.org/ abs/1910.06711

    2019 arXiv

  20. [29]

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

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” 2024. [Online]. Available: https: //arxiv.org/abs/2312.00752

    2024 arXiv

  21. [30]

    Ur- gent challenge: Universality, robustness, and generalizability for speech enhancement,

    W. Zhang, R. Scheibler, K. Saijo, S. Cornell, C. Li, Z. Ni, J. Pirkl- bauer, M. Sach, S. Watanabe, T. Fingscheidt, and Y . Qian, “Ur- gent challenge: Universality, robustness, and generalizability for speech enhancement,” inInterspeech 2024, 2024, pp. 4868–4872

    2024

  22. [31]

    Interspeech 2025 URGENT speech enhancement challenge,

    K. Saijo, W. Zhang, S. Cornell, R. Scheibler, C. Li, Z. Ni, A. Ku- mar, M. Sach, Y . Fu, W. Wang, T. Fingscheidt, and S. Watanabe, “Interspeech 2025 URGENT speech enhancement challenge,”Ac- cepted by Interspeech, 2025

    2025

  23. [32]

    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

  24. [33]

    Icassp 2022 deep noise suppression challenge,

    H. Dubey, V . Gopal, R. Cutler, A. Aazami, S. Matusevych, S. Braun, S. E. Eskimez, M. Thakker, T. Yoshioka, H. Gamper, and R. Aichner, “Icassp 2022 deep noise suppression challenge,”

    2022

  25. [35]

    Nisqa: A deep cnn-self-attention model for multidimensional speech quality pre- diction with crowdsourced datasets,

    G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “Nisqa: A deep cnn-self-attention model for multidimensional speech quality pre- diction with crowdsourced datasets,” in Interspeech 2021, 2021, pp. 2127–2131

    2021

  26. [36]

    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,” in 2001 IEEE In- ternational Conference on Acoustics, Speech, and Signal Process- ing. Proceedings (...

    2001

  27. [38]

    Distance measures for speech process- ing,

    A. Gray and J. Markel, “Distance measures for speech process- ing,” IEEE Transactions on Acoustics, Speech, and Signal Pro- cessing, vol. 24, no. 5, pp. 380–391, 1976

    1976

  28. [39]

    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,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 24, no. 11, pp. 2009–2022, 2016

    2009

  29. [40]

    Evaluation metrics for generative speech en- hancement methods: Issues and perspectives,

    J. Pirklbauer, M. Sach, K. Fluyt, W. Tirry, W. Wardah, S. Moeller, and T. Fingscheidt, “Evaluation metrics for generative speech en- hancement methods: Issues and perspectives,” inSpeech Commu- nication; 15th ITG Conference, 2023, pp. 265–269

    2023

  30. [2022]

    Available: https://arxiv.org/abs/2202.13288

    [Online]. Available: https://arxiv.org/abs/2202.13288

    arXiv

  31. [2023]

    Available: https://arxiv.org/abs/2302.11832

    [Online]. Available: https://arxiv.org/abs/2302.11832

    arXiv

  32. [2024]

    Available: https://arxiv.org/abs/2406.12194

    [Online]. Available: https://arxiv.org/abs/2406.12194

    arXiv

Pith tools

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