Pith. sign in

REVIEW 3 major objections 5 minor 31 references

This paper claims that inserting dual-path recurrent blocks into DeepFilterNet2's encoder, alongside an over-attenuation loss and long-context fine-tuning, lifts a causal speech enhancer past larger open-source models on low-SNR multilingua

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 15:30 UTC pith:L7VGX6KF

load-bearing objection Useful incremental extension of DeepFilterNet2 with released code, but the headline comparison leans on an unreleased eval set and an author-defined relative metric; the gains are plausible, not independently verified. the 3 major comments →

arxiv 2512.16420 v3 pith:L7VGX6KF submitted 2025-12-18 cs.SD

DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN

Daniel Rika , Nino Sapir , Ido Gus This is my paper
classification cs.SD
keywords speech enhancementdual-path RNNcausal streamingDeepFilterNet2over-attenuation lossPRISM metricmultilingual low-SNR evaluationedge NPU deployment
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 paper aims to show that a compact causal speech-enhancement model can outrank much larger open-source systems by upgrading the efficient DeepFilterNet2 architecture with dual-path RNN blocks that capture long-range time and cross-band structure, and by adding two training interventions: an over-attenuation loss that penalises under-estimated speech energy, and a fine-tuning stage on long continuous audio that stabilises always-on streaming. On a newly built 324-clip, 12-language, low-SNR test set, DPDFNet variants place above every causal baseline, including models with ten to twenty times more parameters, and their PRISM aggregate improves monotonically with the number of dual-path blocks (baseline 0.79, DPDFNet-2 0.95, DPDFNet-4 0.98, DPDFNet-8 1.00). The paper also introduces PRISM, a composite metric that min-max normalises intrusive and non-intrusive scores within the evaluated set, and reports that the second-largest variant runs in real time on a commercial edge NPU. If the claim holds, streaming devices with tight power budgets can reach quality previously associated with offline models.

Core claim

On the paper's own terms, the central discovery is that inserting N causal dual-path RNN blocks into both encoder branches of DeepFilterNet2 — after the convolutional stacks that process ERB and complex spectral features — produces a family DPDFNet-{2,4,8} whose performance on the authors' multilingual low-SNR evaluation set improves monotonically with N: DPDFNet-8 reaches PRISM 1.00 and DPDFNet-4 0.98, versus 0.87 for the best non-DPDFNet baseline and 0.75 for DeepFilterNet2. Two training-side components carry much of the gain: an over-attenuation loss that penalises time-frequency bins where the enhanced magnitude falls below the clean target, and a fine-tuning phase on 30-40 second segmen

What carries the argument

The central object is the causal dual-path RNN (DPRNN) block: an intra-stage RNN that scans frequency bins within each time frame (bidirectional) and an inter-stage unidirectional RNN that scans time at each frequency bin, with weights shared across bins and hidden states kept per-bin. Inserted after the convolutional stacks of both encoder branches, it supplies long-range temporal and cross-band context at modest parameter cost (2.49M to 3.54M parameters for 2 to 8 blocks). Two additional mechanisms support the argument: the over-attenuation loss, a binary-masked multi-resolution STFT term that only penalises energy under-shoot relative to the clean target, and the long-segment fine-tuning

Load-bearing premise

The load-bearing premise is the author's own evaluation: a 324-clip, 12-language, low-SNR test set that is not publicly released, and a PRISM metric that min-max normalizes scores within the exact set of models being ranked; if either favors DPDFNet's training pipeline, the claimed consistent gains could vanish on an independent benchmark.

What would settle it

Run the released DPDFNet weights against FastEnhancer-L and DPCRN on an independent public benchmark such as the DNS4 blind test set, and compute PRISM with a fixed reference anchor (e.g., the unprocessed noisy signal for normalisation) instead of min-max across the tested models; if DPDFNet-8 does not rank first or the monotonic depth trend disappears, the claimed superiority is an artifact of the private test set or the relative normalisation.

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

If this is right

  • If the dual-path stack is the cause, a single architecture family scales from about 2.5M to 3.5M parameters with monotonically better PRISM (0.95 to 1.00), so one codebase can serve devices with differing compute budgets by picking N = 2, 4, or 8.
  • The over-attenuation loss and long-context fine-tuning transfer directly to the unchanged DeepFilterNet2-style baseline, raising it from 0.79 to 0.91 on PRISM and ahead of DeepFilterNet3 on most metrics, meaning training fixes alone capture a large share of the improvement.
  • A causal 2.84M-parameter model can beat 33-46M-parameter waveform models on a hard low-SNR multilingual test set, so latency and memory constraints do not necessarily trade away quality.
  • The real-time result on an embedded edge NPU (RT factor 0.97 for the second-largest variant, with int8 weights / int16 activations) indicates the quality gains are achievable under power and latency budgets typical of earbuds and smartwatches, not just server GPUs.

Where Pith is reading between the lines

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

  • A testable extension is pushing the dual-path stack past 8 blocks (12 or 16); the paper's own numbers show a shrinking gain between 4 and 8 blocks, so the monotonic claim likely has a practical depth ceiling worth mapping.
  • Because PRISM min-max normalises within the evaluated model set, PRISM values from different papers are not directly comparable; a public benchmark with a fixed reference anchor would turn PRISM into a portable metric.
  • The curated low-SNR multilingual set — 12 languages, 9 noise types, multi-speaker clips with speech-free gaps up to 15 seconds — is a realistic streaming stress-test that standard English-only benchmarks omit; releasing it would let the field compare causal enhancers under matched conditions.
  • The over-attenuation loss is an asymmetric penalty that fires only when enhanced energy falls below the clean target; that idea could generalise to any mask-based enhancer as a cheap way to curb oversuppression, independent of dual-path architectures.

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

3 major / 5 minor

Summary. The paper proposes DPDFNet, an extension of DeepFilterNet2 that inserts causal dual-path RNN blocks into the encoder branches to strengthen temporal and cross-band modeling. Two training additions are also introduced: an over-attenuation loss that penalizes under-estimation of clean speech, and a fine-tuning stage on 30–40 s continuous segments intended to stabilize an 'always-on' streaming setting. The main evidence is a new, author-curated 324-clip multilingual low-SNR evaluation set, a proposed composite metric PRISM based on min-max normalization of intrusive and non-intrusive metrics, DNSMOS results on the DNS4 blind test set, and real-time factor measurements on Ceva-NeuPro-Nano NPUs. The authors report that DPDFNet variants outperform the selected causal open-source baselines and that performance improves monotonically with the number of DPRNN blocks.

Significance. If the empirical claims are substantiated, the paper is a useful incremental contribution to streaming causal speech enhancement: it combines an established efficient backbone with a known dual-path modeling idea, adds a sensible over-attenuation loss, and demonstrates deployment on an actual edge NPU. The release of code and pretrained models (stated in the abstract/footnote) is a clear strength, as is the on-device runtime evaluation, which is more realistic than CPU-only comparisons. However, the central comparative claim rests on an unreleased, author-curated test set and a metric that is min-max normalized over the very models being ranked. The abstract promises VoiceBank+DEMAND results, but no such results are presented. The DNS4 table only reports DNSMOS and compares only DeepFilterNet2/3 and the proposed models, not the broader baseline set. These issues make the headline 'consistently outperform' claim currently unverifiable by independent readers, despite the released code.

major comments (3)
  1. [Section III-B and Table I] The primary evidence for 'consistently outperform strong causal open-source baselines' is generated on a 324-clip multilingual low-SNR set that is not publicly released. No confidence intervals, significance tests, or per-condition breakdowns are provided, and the table aggregates a single 2.5-minute clip per language/environment/SNR combination. Because the test set is curated by the same authors who trained the models, the comparison is not independently reproducible. The abstract explicitly promises evaluation on VoiceBank+DEMAND and DNS4 'blind test benchmarks,' but VoiceBank+DEMAND results are absent and Table II reports only DNSMOS and only versus DeepFilterNet2/3. Please release the evaluation set (or a public equivalent) and add standard-benchmark tables, including PESQ/STOI/SI-SNR with significance or confidence intervals.
  2. [Section IV-C, Eq. (10) and Table I] PRISM is defined by min-max normalization across the models being ranked, so its scale and all resulting scores depend on which models are included. In Table I, DPDFNet-8 achieves the highest or tied-highest value on nearly every raw metric, so its PRISM of 1.00 is a direct consequence of the normalization range; the metric is relative, not an absolute quality measure. Moreover, no uncertainty or sensitivity analysis is reported, so the 'clear scalability with the number of dual-path blocks' claim is not supported beyond the raw metric trends. The authors should either drop PRISM from the central claims or supplement it with raw-metric confidence intervals, a sensitivity analysis over the normalization set, and a clear statement of what PRISM differences are perceptually meaningful.
  3. [Section IV-B vs Section III-B] The fine-tuning stage uses continuous 30–40 s segments with long silences to improve stability, while the new evaluation set deliberately contains 2.5-minute clips with speech-free intervals up to 15 s. This creates a distributional match between the DPDFNet fine-tuning data and the evaluation set that is not offered to the baseline models (e.g., DTLN, GTCRN, FastEnhancer), which are used with their official weights trained on short segments. The result is a confound: the reported gains may reflect adaptation to the eval set's long-form, silence-heavy structure rather than general enhancement ability. At minimum, the authors should report the baselines' performance after applying the same long-context fine-tuning, and should show results on a standard short-clip benchmark to verify that the fine-tuning does not sacrifice performance on typical conditions.
minor comments (5)
  1. [Abstract] The abstract states 'We evaluate DPDFNet on the standard VoiceBank+DEMAND and DNS4 blind test benchmarks' but the manuscript contains no VoiceBank+DEMAND results. Either add the missing table or revise the abstract.
  2. [Section II-C and Fig. 1] The DPRNN block description says the intra stage uses a bidirectional RNN along the frequency axis, which is non-causal. The text later says the model is causal. Please clarify: if the intra stage is bidirectional, the overall model is not strictly causal in the frequency dimension, or the causality statement needs qualification.
  3. [Section III-C, Eq. (10)] Typographical and grammatical issues: 'both of the lost functions' should be 'both loss functions'; 'an holistic' should be 'a holistic'; 'a pick value' should be 'a peak value'; 'V orbis' presumably should be 'Vorbis'.
  4. [References] Reference [2] attributes DTLN to 'N. Westhausen and N. Zeghidour,' but the author list in the actual paper is Westhausen and Meyer. Please correct the citation and ensure all references are accurate.
  5. [Table I] Several rows show missing spaces in numeric entries (e.g., '3.194.09', '4.154.494.27'), making the table hard to read. Please fix formatting.

Circularity Check

0 steps flagged

No significant circularity: PRISM is a rescaling of external metrics, depth gains are visible in raw scores, and DNS4 provides independent grounding.

full rationale

The paper's derivation chain is not circular. The architectural contribution (dual-path RNN blocks in the DeepFilterNet2 encoder, an over-attenuation loss, and a long-context fine-tuning phase) is evaluated with standard external metrics (PESQ, STOI, SI-SNR, DNSMOS P.835, P.808 MOS, NISQA) on a private multilingual low-SNR set and on the DNS4 blind test set. PRISM is a min-max normalized composite of those external metrics; min-max normalization is a monotonic rescaling that does not change the underlying measured values or rankings. The claimed monotonic improvement with dual-path depth is visible in the raw metrics themselves (e.g., PESQ 3.14→3.18→3.20, STOI 92.6→93.0→93.4, SI-SNR 13.72→14.11→14.47 for DPDFNet-2/4/8), so PRISM is a summary of the evidence, not an input that forces the result. No parameter is fitted to the evaluation set; the fine-tuning stage uses 30–40 sec training segments and is not trained on the test clips. The DNS4 table provides an external benchmark where DPDFNet also improves over DeepFilterNet2/3. The main concerns—absence of the promised VoiceBank+DEMAND results, the unreleased private evaluation set, and the relative nature of PRISM—are reproducibility and benchmark-validity issues, not circularity. There is no load-bearing self-citation or definitional equivalence between the claimed predictions and the inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central empirical claims rest on several hand-chosen hyperparameters (loss weights, depths, schedules), on unvalidated domain assumptions about the reliability of objective metrics, and on a self-created test set plus self-created PRISM metric. The architecture is a combination of cited prior components; the only truly new entity is PRISM, which lacks independent evidence.

free parameters (5)
  • OA loss weight = 500
    Eq. (10) sets lambda_MR = lambda_OA = 500; no ablation over this weight is shown, so DPDFNet's behavior depends on this hand-chosen balance.
  • OA mask hard threshold = M(k,f)=1{|S|>|Y|}
    Eq. (8) uses a strict, no-margin binary mask comparing enhanced and clean magnitudes; the strictness is arbitrary and not swept.
  • DPRNN depth k = 2, 4, 8
    Model variants are defined by these chosen depths; k=8 is reported best while k=4 is real-time, but no architecture search or scaling law justifies the choice.
  • Fine-tuning schedule = 5000 steps, lr 1e-5, 30-40 s segments, batch 1
    Selected ad hoc to address stateful instability; no ablation and no direct stability metric is reported.
  • PRISM normalization set = min-max across the models in Table I on the author test set
    PRISM values are relative to whichever models enter the table; changing the model set would change all scores, making the reported 0.95-1.00 values non-absolute.
axioms (5)
  • domain assumption DeepFilterNet2's two-stage masking/reconstruction (ERB gains + deep-filter coefficients) is an adequate, sufficiently expressive base framework
    Taken from [12] and used as the foundation; the paper does not re-derive its optimality or robustness.
  • domain assumption Dual-path RNN inter/intra modeling improves long-range temporal and cross-band modeling in this setting
    Adopted from [14] and DPCRN [8]; the paper does not prove this holds inside the DeepFilterNet2 encoder, it assumes it.
  • domain assumption PESQ, STOI, SI-SNR, DNSMOS, and NISQA are valid proxies for perceived speech quality
    All quality conclusions are metric-based; no listening test is performed.
  • ad hoc to paper The new 324-clip multilingual low-SNR test set is representative of real-world conditions
    Curated by the authors from Speech-MASSIVE, not publicly released, and used as the primary benchmark; representativeness is asserted, not externally validated.
  • ad hoc to paper PRISM's hierarchical min-max aggregation is a meaningful holistic quality measure
    PRISM is defined in this paper and used to rank models; no independent validation or correlation with subjective scores is shown.
invented entities (1)
  • PRISM composite metric no independent evidence
    purpose: Provides a single scale-normalized score aggregating intrusive and non-intrusive metrics; used to show monotonic improvement with dual-path depth
    Defined in Section IV-C; min-max normalized across the models evaluated on the authors' own test set; no external validation, listening-test correlation, or public standardization is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 11706 in / 10441 out tokens · 103287 ms · 2026-08-03T15:30:26.803941+00:00 · methodology

0 comments
read the original abstract

We present DPDFNet, a causal single-channel speech enhancement model that extends DeepFilterNet2 architecture with dual-path blocks in the encoder, strengthening long-range temporal and cross-band modeling while preserving the original enhancement framework. In addition, we demonstrate that adding a loss component to mitigate over-attenuation in the enhanced speech, combined with a fine-tuning phase tailored for "always-on" applications, leads to substantial improvements in overall model performance. We evaluate DPDFNet on the standard VoiceBank+DEMAND and DNS4 blind test benchmarks, where it shows consistent gains over DeepFilterNet2 and strong overall performance against other causal open-source models. In addition, we introduce a supplementary multilingual low-SNR evaluation set comprising long recordings in 12 languages across everyday noise scenarios, on which DPDFNet delivers superior performance to other causal open-source models, including some that are substantially larger and more computationally demanding. We also propose an holistic metric named PRISM, a composite, scale-normalized aggregate of intrusive and non-intrusive metrics, which demonstrates clear scalability with the number of dual-path blocks. We further demonstrate on-device feasibility by deploying DPDFNet on Ceva-NeuPro-Nano edge NPUs. Results indicate that DPDFNet-4, our second-largest model, achieves real-time performance on NPN32 and runs even faster on NPN64, confirming that state-of-the-art quality can be sustained within strict embedded power and latency constraints.

Figures

Figures reproduced from arXiv: 2512.16420 by Daniel Rika, Ido Gus, Nino Sapir.

Figure 1
Figure 1. Figure 1: Overview of the proposed DPDFNet architecture. (a) single DPRNN block; (b) DeepFilterNet2 scheme integrated with DPRNN blocks. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Impact of dual-path block depth in DPDFNet on PRISM perfor [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Real-time factor of our proposed models on NPN32 and NPN64, assuming deployment with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

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

31 extracted references · 15 linked inside Pith

  1. [1]

    A hybrid dsp/deep learning approach to real-time full-band speech enhancement,

    J.-M. Valin, “A hybrid dsp/deep learning approach to real-time full-band speech enhancement,” 2018. [Online]. Available: https: //arxiv.org/abs/1709.08243

    Pith/arXiv arXiv 2018

  2. [2]

    Dual-signal transformation lstm network for real-time noise suppression,

    N. Westhausen and N. Zeghidour, “Dual-signal transformation lstm network for real-time noise suppression,” inProc. Interspeech, 2020, pp. 2472–2476. [Online]. Available: https://www.isca-speech.org/archive/ interspeech 2020/westhausen20 interspeech.html

    2020

  3. [3]

    Towards efficient models for real-time deep noise suppression,

    S. Braun, H. Gamper, C. K. A. Reddy, and I. Tashev, “Towards efficient models for real-time deep noise suppression,” 2021. [Online]. Available: https://arxiv.org/abs/2101.09249

    Pith/arXiv arXiv 2021

  4. [4]

    Real time speech enhancement in the waveform domain,

    A. D ´efossez, G. Synnaeve, and Y . Adi, “Real time speech enhancement in the waveform domain,” 2020. [Online]. Available: https://arxiv.org/abs/2006.12847

    Pith/arXiv arXiv 2020

  5. [5]

    Speech denoising in the waveform domain with self-attention,

    Z. Kong, W. Ping, A. Dantrey, and B. Catanzaro, “Speech denoising in the waveform domain with self-attention,” 2022. [Online]. Available: https://arxiv.org/abs/2202.07790

    Pith/arXiv arXiv 2022

  6. [6]

    Fullsubnet: A full-band and sub- band fusion model for real-time single-channel speech enhancement,

    X. Hao, X. Su, R. Horaud, and X. Li, “Fullsubnet: A full-band and sub- band fusion model for real-time single-channel speech enhancement,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, jun 2021. [Online]. Available: http://dx.doi.org/10.1109/ICASSP39728.2021.9414177 8

    arXiv 2021

  7. [7]

    Complex ratio masking for monaural speech separation,

    D. S. Williamson, Y . Wang, and D. Wang, “Complex ratio masking for monaural speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 24, no. 3, pp. 483–492, 2016

    2016

  8. [8]

    Dpcrn: Dual-path convolution recurrent network for single channel speech enhancement,

    X. Le, H. Chen, K. Chen, and J. Lu, “Dpcrn: Dual-path convolution recurrent network for single channel speech enhancement,” inProc. Interspeech, 2021, pp. 2811–2815. [Online]. Available: https://www. isca-speech.org/archive/interspeech 2021/le21b interspeech.html

    2021

  9. [9]

    Gtcrn: A speech enhancement model requiring ultralow computational resources,

    X. Rong, T. Sun, X. Zhang, Y . Hu, C. Zhu, and J. Lu, “Gtcrn: A speech enhancement model requiring ultralow computational resources,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 971–975

    2024

  10. [10]

    Fastenhancer: Speed- optimized streaming neural speech enhancement,

    S. Ahn, J. Han, B. J. Woo, and N. S. Kim, “Fastenhancer: Speed- optimized streaming neural speech enhancement,” 2025. [Online]. Available: https://arxiv.org/abs/2509.21867

    arXiv 2025

  11. [11]

    atennuate: Optimized real-time speech enhancement with deep ssms on raw audio,

    Y . R. Pei, R. Shrivastava, and F. Sidharth, “atennuate: Optimized real-time speech enhancement with deep ssms on raw audio,” 2025. [Online]. Available: https://arxiv.org/abs/2409.03377

    Pith/arXiv arXiv 2025

  12. [12]

    Deepfilternet2: Towards real-time speech enhancement on embedded devices for full-band audio,

    H. Schr ¨oter, A. N. Escalante-B., T. Rosenkranz, and A. Maier, “Deepfilternet2: Towards real-time speech enhancement on embedded devices for full-band audio,” 2022. [Online]. Available: https: //arxiv.org/abs/2205.05474

    Pith/arXiv arXiv 2022

  13. [13]

    Speech- MASSIVE: A Multilingual Speech Dataset for SLU and Beyond,

    B. Lee, I. Calapodescu, M. Gaido, M. Negri, and L. Besacier, “Speech- MASSIVE: A Multilingual Speech Dataset for SLU and Beyond,” in Proc. Interspeech 2024, 2024

    2024

  14. [14]

    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,”

  15. [15]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” 2016. [Online]. Available: https://arxiv.org/abs/1607.06450

    Pith/arXiv arXiv 2016

  16. [16]

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

    Pith/arXiv arXiv 2022

  17. [17]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” in Interspeech 2020. ISCA, Oct. 2020. [Online]. Available: http: //dx.doi.org/10.21437/Interspeech.2020-2826

    doi:10.21437/interspeech.2020-2826 2020

  18. [18]

    Musan: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech, and noise corpus,” 2015. [Online]. Available: https://arxiv.org/abs/1510.08484

    Pith/arXiv arXiv 2015

  19. [19]

    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,” 2022. [Online]. Available: https://arxiv.org/abs/2010.00475

    Pith/arXiv arXiv 2022

  20. [20]

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

    C. Valentini-Botinhao, X. Wang, S. Takaki, and J. Yamagishi, “Investi- gating 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

  21. [21]

    Deepfilternet: Perceptually motivated real-time speech enhancement,

    H. Schr ¨oter, T. Rosenkranz, A. N. Escalante-B., and A. Maier, “Deepfilternet: Perceptually motivated real-time speech enhancement,”

  22. [22]

    Perceptual evaluation of speech quality (pesq) – a new method for speech quality assessment of telephone networks and codecs,

    A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hekstra, “Perceptual evaluation of speech quality (pesq) – a new method for speech quality assessment of telephone networks and codecs,” inProc. IEEE ICASSP, 2001, pp. 749–752

    2001

  23. [23]

    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. [Online]. Available: https://sps.ewi.tudelft.nl/pubs/Taal2011 1.pdf

    2011

  24. [24]

    Sdr – half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “Sdr – half-baked or well done?” inProc. IEEE ICASSP, 2019, pp. 626–630. [Online]. Available: https://www.jonathanleroux.org/pdf/ LeRoux2019ICASSP05sdr.pdf

    2019

  25. [25]

    Dnsmos p.835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,

    C. K. A. Reddy, V . Gopal, and R. Cutler, “Dnsmos p.835: A non- intrusive perceptual objective speech quality metric to evaluate noise suppressors,” 2022. [Online]. Available: https://arxiv.org/abs/2110.01763

    Pith/arXiv arXiv 2022

  26. [26]

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

    ——, “Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors,” 2021. [Online]. Available: https://arxiv.org/abs/2010.15258

    Pith/arXiv arXiv 2021

  27. [27]

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

    G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “Nisqa: A deep cnn-self-attention model for multidimensional speech quality prediction with crowdsourced datasets,” inInterspeech 2021. ISCA, Aug. 2021. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2021-299

    doi:10.21437/interspeech.2021-299 2021

  28. [28]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,”

  29. [2019]

    Available: https://arxiv.org/abs/1711.05101

    [Online]. Available: https://arxiv.org/abs/1711.05101

    Pith/arXiv arXiv

  30. [2020]

    Available: https://arxiv.org/abs/1910.06379

    [Online]. Available: https://arxiv.org/abs/1910.06379

    Pith/arXiv arXiv 1910

  31. [2023]

    Available: https://arxiv.org/abs/2305.08227

    [Online]. Available: https://arxiv.org/abs/2305.08227

    Pith/arXiv arXiv