Pith. sign in

REVIEW 5 major objections 3 minor 43 references

A High-Fidelity Speech Super Resolution Network using a Complex Global Attention Module with Spectro-Temporal Loss

T0 review · 5 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A fully complex U-Net with global time-frequency attention claims to reconstruct speech magnitude and phase jointly, beating four state-of-the-art bandwidth-extension models on VCTK, with its largest wins at 2-to-48 kHz upsampling.

desk verdict A clever architecture undermined by a wrong core equation and inconsistent tables; the idea has merit, but the paper as submitted doesn't hold together. read the letter →

arxiv 2507.00229 v1 pith:D4G5N7JK submitted 2025-06-30 cs.SD cs.AIeess.AS

Tarikul Islam Tamiti , Biraj Joshi , Rida Hasan , Rashedul Hasan , Taieba Athay , Nursad Mamun , Anomadarshi Barua This is my paper · ORCID
classification cs.SDcs.AIeess.AS
keywords speechsuper-resolutionbandwidthextensioncomplex-valuedneuralnetworkcomplexglobalattentionphasereconstructionmulti-resolutionSTFTlossSI-SDRVCTK
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

CTFT-Net is a speech bandwidth-extension model that works entirely on complex-valued time-frequency spectrograms, so it predicts phase as well as magnitude instead of discarding phase or relying on a vocoder. The paper claims that on the VCTK corpus it beats four published super-resolution systems—NU-Wave, WSRGlow, NVSR, and AERO—on log-spectral distance, with the largest gains at extreme upsampling from 2 kHz to 48 kHz. The mechanism is a complex global attention block that models correlations along both the time and frequency axes, a complex conformer in the bottleneck, and a loss that combines time-domain SI-SDR with a multi-resolution real-part STFT loss. A careful reader would care because high-frequency content and phase drive perceived speech quality, and the result suggests a single complex U-Net can do 24x bandwidth extension without post-processing tricks.

What carries the argument

The load-bearing object is the complex-valued 2D convolution of Eq. (1), which the paper uses to build encoders, decoders, skip blocks, attention blocks, and the conformer; it lets the network treat each spectrogram as a complex tensor rather than handling magnitude and phase separately. Around that operation, the complex global attention block (CGAB) reshapes encoder features along the time and frequency axes, applies separate dense layers to each axis, and recombines them to model inter-phoneme and harmonic correlations. A complex conformer in the bottleneck adds multi-head self-attention and local convolution, and the loss combines real-part multi-resolution STFT terms with SI-SDR so the model is supervised in both frequency and time domains.

What would settle it

Inspect the released training code (the paper says source will be available after acceptance) for the complex-convolution layer. If the imaginary branch computes $W_r \ast S_i - W_i \ast S_r$ as written, retraining with the standard $W_r \ast S_i + W_i \ast S_r$ and comparing LSD on VCTK 2-to-48 kHz upsampling would settle whether the claimed improvement depends on the non-standard transform; if instead the code uses the standard update, then Eq. (1) misstates the architecture and the paper's description of its core mechanism is wrong.

Watch

Extended reading notes

Core claim

The central claim is that magnitude and phase of a low-resolution speech spectrogram can be regenerated together by a U-Net whose every layer is complex-valued. The network takes an STFT of the low-resolution waveform, processes it through eight complex encoders and eight complex decoders with complex skip connections, inserts a complex global attention block after the first and seventh encoders to capture inter-phoneme and inter-frequency correlations, and places a complex conformer in the bottleneck to mix local and global context. It is trained by summing a real multi-resolution STFT loss (spectral convergence plus log-magnitude on the real part, at three STFT resolutions) with a time-domain SI-SDR loss. On VCTK the paper reports the lowest LSD among the four compared systems at 2, 4, 8, and 12 kHz input to a 48 kHz target, for example 1.06 versus 1.10 for NVSR at 2 kHz and 0.62 versus 0.87 at 12 kHz, while leaving SI-SDR essentially flat, which the authors read as absence of added noise. The paper also claims this removes the need for a vocoder, for copying input low-frequency bands at the output, and for flipping or reusing the input phase.

Load-bearing premise

The load-bearing premise is that Eq. (1) as printed, with the imaginary update described as real kernel times imaginary input minus imaginary kernel times real input, is what the network actually computes; if it is a typo, the paper misstates its central building block, and if it is not, the network applies a non-standard transform.

Editorial extensions

If this is right

  • A vocoder-free, post-processing-free bandwidth-extension pipeline that predicts phase directly could replace vocoder-dependent systems such as NVSR.
  • Extreme upsampling ratios, from telephony-grade 2 kHz to 48 kHz, become feasible with a single network, with the reported LSD dropping from 1.06 at 2 kHz input to 0.62 at 12 kHz input.
  • Joint training with multi-resolution STFT and SI-SDR improves perceived quality, as measured by PESQ, while keeping intelligibility and SI-SDR nearly unchanged, indicating no noisy artifacts at the band edge.
  • The complex global attention block, which attends to both axes of a complex spectrogram, is the component that the ablation study credits for the main gains over frequency-only attention.

Reading between the lines

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

  • If Eq. (1)'s imaginary update is what was actually implemented, the network is not performing standard complex convolution; retraining with the standard update $W_r \ast S_i + W_i \ast S_r$ would test whether the reported gains persist, because the paper does not clarify whether the equation or the code is authoritative.
  • The CGAB's dual-axis global attention on complex spectrograms is a natural component to reuse in speech enhancement, dereverberation, or text-to-speech systems that need high-frequency and phase fidelity, although the paper does not test those tasks.
  • The paper only evaluates fixed input resolutions on speech from VCTK; applying the same loss and attention design to variable bandwidths or to music would show whether the mechanism generalizes beyond the reported setting.
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

5 major / 3 minor

Summary. The paper proposes CTFT-Net, a complex-valued U-Net for speech super-resolution that takes a complex STFT input, processes it through complex encoders/decoders, complex skip blocks, a complex global attention block (CGAB), and a complex conformer bottleneck, and is trained with a combination of time-domain SI-SDR loss and multiresolution STFT loss. The authors report experiments on the VCTK dataset with target 48 kHz and 16 kHz sampling rates and claim that CTFT-Net outperforms NU-Wave, WSRGlow, NVSR, and AERO, particularly for extreme 2 kHz to 48 kHz upsampling, while jointly recovering magnitude and phase without a vocoder.

Significance. If the claims were substantiated, the paper would make a useful engineering contribution: a single complex-domain network that reconstructs both magnitude and phase for bandwidth extension, with an attention mechanism spanning time and frequency axes, would be attractive compared with vocoder-based pipelines. The ablation study is also a good-faith attempt to isolate the effects of CGAB placement, loss choices, and bottleneck architecture. However, the empirical evidence as presented is not reliable: the central performance claim is contradicted by the paper's own table, the core complex-convolution definition is wrong as written, and the ablation table contains impossible metric values. There is no released code, no checkpoints, and no uncertainty quantification, so the claimed improvements cannot be independently verified. The strengths of the paper are conceptual and architectural, not empirical at this stage.

major comments (5)
  1. [4.1, Table 1] The central claim that CTFT-Net achieves the lowest LSD among all baselines is contradicted by the numbers in Table 1. For the 2 kHz input, the proposed method reports LSD 1.06, while AP-BWE (also in the table) reports 1.016; for the 4 kHz input, the proposed method reports 0.96, while AP-BWE reports 0.92. The text states that "for each LR input, CTFT-Net achieves the lowest LSD compared to all baselines," which is not true if AP-BWE is counted as a baseline. The comparison must be restated to name only the baselines for which the claim actually holds, or the table must be corrected if the AP-BWE numbers are in error.
  2. [2.2, Eq. (1)] Equation (1) does not define a valid complex convolution. For a complex kernel W = Wr + jWi and input Sin = Sr + jSi, the standard complex multiplication gives Si0 = Wr * Si_in + Wi * Sr_in, whereas the paper writes Si0 = Wr * Si_in - Wi * Sr_in. This is not a typo in sign convention: it changes the operation from complex convolution to a nonstandard bilinear map. Since the paper explicitly says that every complex layer, including CGAB and the complex conformer, implements Eq. (1), the submitted text either describes a network that is not performing complex convolution or contains a formula error in the load-bearing definition. This must be resolved, ideally by providing code or a corrected derivation.
  3. [4.2, Table 3] Table 3 contains NISQA-MOS values that are outside the valid range of the metric. NISQA-MOS is a mean opinion score on a 1-5 scale, yet rows P7, P8, and P9 report values of 11.47, 11.19, and 11.47, respectively. These values are physically impossible, which indicates a data-entry error or an undefined modification of the metric. Because Table 3 is the basis for the ablation conclusions, all rows must be re-verified and recomputed.
  4. [4.2, Table 3, row P10] Row P10, which removes the SI-SDR loss, reports an LSD of 0.88 — better than the proposed CTFT-Net configurations in P6.1 (1.06) and P6.2 (1.01). The paper claims that combining SI-SDR loss with MR-STFT loss improves audio quality, and the ablation is used to support that claim, but P10 shows the opposite for LSD. The authors must explain this discrepancy, for example by discussing loss weighting or the trade-off between LSD and SI-SDR, or by correcting the table.
  5. [3.3, 4.1] No code, checkpoints, or evaluation scripts are provided, and no error bars, confidence intervals, or per-utterance statistics are reported for any metric. The paper states "Source code of the model will be available after acceptance," which prevents a reviewer or reader from verifying the central empirical claims, especially given the inconsistencies in Tables 1 and 3. Without these artifacts or statistical measures, the claimed superiority over SOTA systems is not reproducible.
minor comments (3)
  1. [2.6, Eq. (2)] Equation (2) has a formatting issue: the summation symbols and parentheses are not properly typeset, making it hard to see that the spectral convergence and log-magnitude losses are summed over resolutions.
  2. [Table 3] Row label P6 is used in the text but the table contains P6.1 and P6.2, making the reference ambiguous; the text should say which configuration is meant.
  3. [5, Conclusion] The conclusion repeats the claim of "strong performance" without acknowledging the discrepancies in Table 1 or the impossibility of some numbers in Table 3; a more cautious summary would be appropriate.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central benchmark claim is external and falsifiable; the only author-overlapping citation is used as non-load-bearing context.

full rationale

CTFT-Net is an empirical engineering contribution whose central claim is benchmark performance against external open-source baselines (NU-Wave, WSRGlow, NVSR, AERO) on VCTK. The architecture is specified by explicit equations and ablations, with no fitted parameter later relabeled as a prediction: the reported LSD, STOI, PESQ, and SI-SDR values are measured on held-out speakers against independently released code, so the main claim is externally falsifiable. The only author-overlapping reference is [25] (Mamun & Hansen), cited in Section 2.4 only to note that T-F attention modules have already been proposed for speech enhancement and dereverberation; the cited work supplies neither a uniqueness theorem nor any premise from which CTFT-Net's design or results follow, so the self-citation is not load-bearing. Section 2.2 Eq. (1) has a sign inconsistency in the imaginary-part update (Wr*Si_in - Wi*Sr_in instead of Wr*Si_in + Wi*Sr_in under the real-part convention); this is a correctness or documentation concern, not a circular reduction, and cannot support a circularity finding. The paper therefore presents no derivation chain in which an output reduces by construction to an input; the low score reflects only the minor context-only self-citation.

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

No free physical constants or invented entities appear. The central empirical claim depends on hand-selected architecture choices and on the correctness of the complex-convolution definition in Eq. (1), which appears mis-stated. The model's 61.6M learned weights are fitted to VCTK data by gradient descent; they are excluded here as ordinary ML training, but evaluation code and weights are not released.

free parameters (4)
  • CGAB placement = 2 blocks (between encoders 1-2 and 7-8)
    Chosen by design trade-off; ablation P2 puts CGAB in every encoder and gives a small LSD improvement (1.06 to 1.04) at 31% more parameters. Not derived from theory.
  • MR-STFT resolutions = S=3 with {(256,128,256), (512,256,512), (1024,512,1024)}
    Hand-selected; only the comparison of MR over single-resolution (P5) is reported, with no resolution search.
  • Loss combination weights = Lr_MR-STFT + SI-SDR, equal weights
    Weights are not specified or tuned; P10 shows removing SI-SDR changes LSD from 1.06 to 0.88, so the chosen combination is not justified by the reported headline metric.
  • Channel schedule = Encoder 1->24->48->48->72->72->96->96->192, with matching decoder sizes
    Manual architecture hyperparameters; no ablation or design rationale is given.
assumptions (4)
  • standard math STFT analysis-synthesis is invertible for the three chosen resolutions
    The model maps waveform to complex STFT and back; this relies on standard STFT reconstruction, not shown in the paper.
  • domain assumption Sixth-order low-pass filtering, downsampling, and sinc interpolation produce valid low-resolution speech
    Section 3.1 describes LR simulation; the claim 'without noisy artifacts' depends on this simulation matching real low-bandwidth recordings.
  • domain assumption The reproduced baselines NU-Wave, WSRGlow, NVSR, and AERO were evaluated under comparable conditions
    Section 4.1 says baselines were reproduced with open-source code and default settings, but no logs or model checkpoints are provided.
  • ad hoc to paper Complex convolution is the operation in Eq. (1) as written
    If corrected to the standard Wr*Si + Wi*Sr, the paper mis-describes its own architecture; if not corrected, the network computes a non-standard transform.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A High-Fidelity Speech Super Resolution Network using a Complex Global Attention Module with Spectro-Temporal Loss." pith.science (2026). https://pith.science/paper/D4G5N7JK

@misc{pith2026250700229,
  author       = {Pith},
  title        = {Pith review of: A High-Fidelity Speech Super Resolution Network using a Complex Global Attention Module with Spectro-Temporal Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4G5N7JK}},
  note         = {Machine review of arXiv:2507.00229}
}
read the original abstract

Speech super-resolution (SSR) enhances low-resolution speech by increasing the sampling rate. While most SSR methods focus on magnitude reconstruction, recent research highlights the importance of phase reconstruction for improved perceptual quality. Therefore, we introduce CTFT-Net, a Complex Time-Frequency Transformation Network that reconstructs both magnitude and phase in complex domains for improved SSR tasks. It incorporates a complex global attention block to model inter-phoneme and inter-frequency dependencies and a complex conformer to capture long-range and local features, improving frequency reconstruction and noise robustness. CTFT-Net employs time-domain and multi-resolution frequency-domain loss functions for better generalization. Experiments show CTFT-Net outperforms state-of-the-art models (NU-Wave, WSRGlow, NVSR, AERO) on the VCTK dataset, particularly for extreme upsampling (2 kHz to 48 kHz), reconstructing high frequencies effectively without noisy artifacts.

Figures

Figures reproduced from arXiv: 2507.00229 by the authors.

Figure 1
Figure 1. CTFT-Net has complex encoders, decoders, complex skip blocks, CGAB, and real MR-STFT + SI-SDR loss. proposed model on the VCTK multispeaker dataset. Results show that CTFT-Net outperforms the SOTA SSR models. 2. Methodology Here, we discuss our proposed modifications on U-Net that con￾struct complex-valued CTFT-Net for SSR tasks. 2.1. Proposed network architecture in complex-domain The detailed architecture of the p… view at source ↗
Figure 2
Figure 2. CGAB captures complex global T-F correlations. Based on this understanding, we propose a self-attention technique along the T-F axes within each spectrogram, with￾out considering correlations among consecutive spectrograms at this stage (see Section 2.5). Specifically, attention on frequency and time axes are implemented by two separate fully connected (FC) layers. Along the time path, the input and output dimen￾sio… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [1]

    Therefore, SSR is making its way into different practical applications, where speech quality enhancement [2] and text-to-speech synthesis [3] are required

    Introduction Speech super-resolution (SSR), also known as bandwidth ex- tension (BWE) [1], generates missing high frequencies from low-frequency speech contents to improve speech clarity and naturalness. Therefore, SSR is making its way into different practical applications, where speech quality enhancement [2] and text-to-speech synthesis [3] are require...

    arXiv 2025

  2. [2]

    Methodology Here, we discuss our proposed modifications on U-Net that con- struct complex-valued CTFT-Net for SSR tasks. 2.1. Proposed network architecture in complex-domain The detailed architecture of the proposed CTFT-Net is shown in Fig. 1. The network consists of four main components: (i) a total of 16 (i.e., 8 + 8) full complex-valued encoder-decode...

    work page

  3. [3]

    Experiments 3.1. Speech corpus and preprocessing We use VCTK (version 0.92) [30], a multi-speaker English cor- pus containing 110 speakers, for training (i.e., 95 speakers) and testing (i.e., 11 speakers). Each audio clip has a duration rang- ing from 2s to 7s. We standardize all audio clips to 4s by either zero-padding or trimming. Following [9], only th...

    work page

  4. [4]

    Results We conduct comprehensive evaluations of CTFT-Net by com- paring it with SOTA models, followed by an ablation study. 4.1. Performance analysis Table 1: LSD Comparison for 48 kHz target sampling rate. Model 2 kHz 4 kHz 8 kHz 12 kHz Size (M) Unprocessed 3.06 2.85 2.44 1.34 - NU-Wave [13] 1.85 1.48 1.45 1.27 3 WSRGlow [14] 1.45 1.18 1.02 0.91 - (40) N...

    work page

  5. [5]

    It shows strong performance across a wide range of input sampling rates ranging from 2 kHz to 48 kHz

    Conclusion This paper presents a novel SSR framework that operates en- tirely in complex domains, jointly reconstructing both magni- tude and phase from the LR signal using global attention on T-F axes. It shows strong performance across a wide range of input sampling rates ranging from 2 kHz to 48 kHz. For the VCTK multi-speaker benchmark, results show t...

    work page

  6. [6]

    Nu-wave 2: A general neural audio up- sampling model for various sampling rates,

    S. Han and J. Lee, “Nu-wave 2: A general neural audio up- sampling model for various sampling rates,” in Interspeech 2022, 2022, pp. 4401–4405

    work page 2022

  7. [7]

    Bandwidth extension is all you need,

    J. Su, Y . Wang, A. Finkelstein, and Z. Jin, “Bandwidth extension is all you need,” in ICASSP 2021-2021 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2021, pp. 696–700

    work page 2021

  8. [8]

    Speech enhancement via frequency band- width extension using line spectral frequencies,

    S. Chennoukh et al., “Speech enhancement via frequency band- width extension using line spectral frequencies,” in 2001 IEEE International Conference on Acoustics, Speech, and Signal Pro- cessing. Proceedings, vol. 1. IEEE, 2001, pp. 665–668

    work page 2001

Show all 43 references
  1. [9]

    A mel-cepstral analysis technique restoring high fre- quency components from low-sampling-rate speech

    K. Nakamura, K. Hashimoto, K. Oura, Y . Nankaku, and K. Tokuda, “A mel-cepstral analysis technique restoring high fre- quency components from low-sampling-rate speech.” in Inter- speech, 2014, pp. 2494–2498

    2014

  2. [10]

    Audio super- resolution with robust speech representation learning of masked autoencoder,

    S.-B. Kim, S.-H. Lee, H.-Y . Choi, and S.-W. Lee, “Audio super- resolution with robust speech representation learning of masked autoencoder,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  3. [11]

    Real-time speech frequency bandwidth extension,

    Y . Li, M. Tagliasacchi, O. Rybakov, V . Ungureanu, and D. Roblek, “Real-time speech frequency bandwidth extension,” in ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 691–695

    2021

  4. [12]

    Phasen: A phase-and- harmonics-aware speech enhancement network,

    D. Yin, C. Luo, Z. Xiong, and W. Zeng, “Phasen: A phase-and- harmonics-aware speech enhancement network,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 05, 2020, pp. 9458–9465

    2020

  5. [13]

    Bandwidth extension of musical au- dio signals with no side information using dilated convolutional neural networks,

    M. Lagrange and F. Gontier, “Bandwidth extension of musical au- dio signals with no side information using dilated convolutional neural networks,” inICASSP 2020-2020 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2020, pp. 801–805

    2020

  6. [14]

    Speech band- width extension using generative adversarial networks,

    S. Li, S. Villette, P. Ramadas, and D. J. Sinder, “Speech band- width extension using generative adversarial networks,” in 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 5029–5033

    2018

  7. [15]

    Nu-gan: High resolution neural upsampling with gan,

    R. Kumar, K. Kumar, V . Anand, Y . Bengio, and A. Courville, “Nu-gan: High resolution neural upsampling with gan,” arXiv preprint arXiv:2010.11362, 2020

    2010 arXiv

  8. [16]

    Speech super resolution gen- erative adversarial network,

    S. E. Eskimez and K. Koishida, “Speech super resolution gen- erative adversarial network,” in ICASSP 2019-2019 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 3717–3721

    2019

  9. [17]

    A deep neural network approach to speech bandwidth expansion,

    K. Li and C.-H. Lee, “A deep neural network approach to speech bandwidth expansion,” in2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015, pp. 4395–4399

    2015

  10. [18]

    Au- diosr: Versatile audio super-resolution at scale,

    H. Liu, K. Chen, Q. Tian, W. Wang, and M. D. Plumbley, “Au- diosr: Versatile audio super-resolution at scale,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 1076–1080

    2024

  11. [19]

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

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

    2021

  12. [20]

    Wsrglow: A glow-based waveform generative model for audio super-resolution,

    K. Zhang, Y . Ren, C. Xu, and Z. Zhao, “Wsrglow: A glow-based waveform generative model for audio super-resolution,” in Inter- speech 2021, 2021, pp. 1649–1653

    2021

  13. [21]

    Neu- ral vocoder is all you need for speech super-resolution,

    H. Liu, W. Choi, X. Liu, Q. Kong, Q. Tian, and D. Wang, “Neu- ral vocoder is all you need for speech super-resolution,” in Inter- speech, 2022

    2022

  14. [22]

    Aero: Audio super resolution in the spectral domain,

    M. Mandel, O. Tal, and Y . Adi, “Aero: Audio super resolution in the spectral domain,” in ICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  15. [23]

    Sdnet: Noise-robust bandwidth extension under flexible sampling rates,

    J. Yang, H. Liu, L. Gan, Y . Zhou, X. Li, J. Jia, and J. Yao, “Sdnet: Noise-robust bandwidth extension under flexible sampling rates,” in 2024 Asia Pacific Signal and Information Processing Associa- tion Annual Summit and Conference (APSIPA ASC). IEEE, 2024, pp. 1–6

    2024

  16. [24]

    Dpt-fsnet: Dual-path trans- former based full-band and sub-band fusion network for speech enhancement,

    F. Dang, H. Chen, and P. Zhang, “Dpt-fsnet: Dual-path trans- former based full-band and sub-band fusion network for speech enhancement,” in ICASSP 2022-2022 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 6857–6861

    2022

  17. [25]

    Sdr– half-baked or well done?

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

    2019

  18. [26]

    Tfgan: Time and frequency domain based generative adversar- ial network for high-fidelity speech synthesis,

    Q. Tian, Y . Chen, Z. Zhang, H. Lu, L. Chen, L. Xie, and S. Liu, “Tfgan: Time and frequency domain based generative adversar- ial network for high-fidelity speech synthesis,” arXiv preprint arXiv:2011.12206, 2020

    2011 arXiv

  19. [27]

    Skipconvnet: Skip convolutional neural network for speech dereverberation using optimally smoothed spectral map- ping,

    V . Kothapally, W. Xia, S. Ghorbani, J. H. Hansen, W. Xue, and J. Huang, “Skipconvnet: Skip convolutional neural network for speech dereverberation using optimally smoothed spectral map- ping,” in Interspeech 2020, 2020, pp. 3935–3939

    2020

  20. [28]

    Joint time- frequency and time domain learning for speech enhancement,

    C. Tang, C. Luo, Z. Zhao, W. Xie, and W. Zeng, “Joint time- frequency and time domain learning for speech enhancement,” in Proceedings of the twenty-ninth international conference on in- ternational joint conferences on artificial intelligence , 2021, pp. 3816–3822

    2021

  21. [29]

    A time-frequency attention module for neural speech en- hancement,

    Q. Zhang, X. Qian, Z. Ni, A. Nicolson, E. Ambikairajah, and H. Li, “A time-frequency attention module for neural speech en- hancement,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 462–475, 2022

    2022

  22. [30]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),

    J. Yamagishi, C. Veaux, K. MacDonald et al., “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (ver- sion 0.92),” University of Edinburgh. The Centre for Speech Tech- nology Research (CSTR), pp. 271–350, 2019

    2019

  23. [31]

    for intelligibility, perceptual evaluation of speech quality (PESQ) [32] for perceived quality, and scale-invariant signal- to-distortion ratio (SI-SDR) [19] for overall signal distortion

  24. [32]

    Speech enhancement for cochlear implant recipients using deep complex convolution transformer with frequency transformation,

    N. Mamun and J. H. Hansen, “Speech enhancement for cochlear implant recipients using deep complex convolution transformer with frequency transformation,” IEEE/ACM Transactions on Au- dio, Speech, and Language Processing, 2024

    2024

  25. [33]

    Complex-valued time-frequency self-attention for speech dereverberation,

    V . Kothapally and J. H. Hansen, “Complex-valued time-frequency self-attention for speech dereverberation,” in Interspeech, 2022

    2022

  26. [34]

    Attention is all you need,

    V . Ashish, “Attention is all you need,” Advances in neural infor- mation processing systems, vol. 30, p. I, 2017

    2017

  27. [35]

    Speech enhancement based on a priori signal to noise estimation,

    P. Scalart et al., “Speech enhancement based on a priori signal to noise estimation,” in 1996 IEEE International Conference on Acoustics, Speech, and Signal Processing Conference Proceed- ings, vol. 2. IEEE, 1996, pp. 629–632

    1996

  28. [36]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in Interspeech 2020, 2020, pp. 5036–5040

    2020

  29. [37]

    An al- gorithm for intelligibility prediction of time–frequency weighted noisy speech,

    C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen, “An al- gorithm 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

    2011

  30. [38]

    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,” in 2001 IEEE international conference on acoustics, speech, and signal processing. Proceed...

    2001

  31. [39]

    Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction,

    Y .-X. Lu, Y . Ai, H.-P. Du, and Z.-H. Ling, “Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  32. [40]

    Nu-wave - official pytorch implementation,

    “Nu-wave - official pytorch implementation,” https://github.com/ maum-ai/nuwave, accessed: 2025-02-16

    2025

  33. [41]

    “Nvsr,” https://github.com/haoheliu/ssr_eval, accessed: 2025-02- 16

    2025

  34. [42]

    “Aero,” https://github.com/slp-rl/aero, accessed: 2025-02-16

    2025

  35. [43]

    Wsrglow,

    “Wsrglow,” https://github.com/zkx06111/WSRGlow, accessed: 2025-02-16

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.