Pith. sign in

REVIEW 4 major objections 5 minor 57 references

Efficient Neural and Numerical Methods for High-Quality Online Speech Spectrogram Inversion via Gradient Theorem

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read An 8,464-parameter CNN and a tridiagonal solver make online speech spectrogram inversion efficient with no loss in measured quality.

desk verdict Solid engineering paper with a genuinely useful tridiagonal observation, but the quality-maintenance claim hinges on unspecified weighting matrices from the prior work. read the letter →

arxiv 2505.24498 v1 pith:MQWZWKWD submitted 2025-05-30 cs.LG

classification cs.LG
keywords spectrograminversiononlinephasereconstructiongradienttheoremdeeplearningderivativestridiagonallinearsystemscausalconvolutionalnetworksreal-timespeechprocessing
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

Reconstructing a speech waveform from its spectrogram in real time has usually meant trading away quality or compute. This paper claims that the two-stage framework of the prior state of the art can be made drastically cheaper without a measured drop in quality, through three changes: a causal CNN with only 8.46k parameters, a strided inference mode that halves the network's cost in exchange for one hop of latency, and the observation that the least-squares phase-recovery step is a tridiagonal positive-semidefinite linear system solvable in linear time and memory. The authors benchmark the resulting pipeline on the LibriSpeech dataset and report ESTOI, WB-PESQ, and log-spectral convergence scores close to ground-truth phase reconstruction and statistically comparable to their implementation of the prior method. If these results hold across other window, hop, and spectrogram settings, real-time spectrogram inversion becomes a realistic option for low-power and interactive speech applications.

What carries the argument

The load-bearing object is the second-stage linear system from Equation 11, A z = b with A = Λ_{τ0} + D_{τ0}^H Γ_{τ0} D_{τ0}. Expanding the Gram term shows that D^H Γ D contributes only diagonal, subdiagonal, and superdiagonal entries (Equation 12), because D is a bidiagonal matrix with -u on the main diagonal and ones on the superdiagonal, and Λ, Γ are diagonal nonnegative weighting matrices. Tridiagonality plus positive semidefiniteness lets the authors solve the system with Thomas' algorithm in O(L) time and memory, and Equation 12 gives a direct recipe for constructing the three diagonals without forming A explicitly. The other key component is the 8.46k-parameter causal CNN designed to predict the FPD and BPD features with a stem-body-head architecture, concatenation instead of residuals, batch normalization, leaky ReLUs, and 1×1 convolutions; its joint output head and optional stride-2 temporal downsampling realize the parameter and computation reductions.

What would settle it

Re-run the evaluation of the proposed 8.46k-parameter CNN plus tridiagonal solver against the original published model and code of the prior method (not the authors' reimplementation) on the same LibriSpeech test-clean conditions; a statistically significant drop in ESTOI or WB-PESQ would contradict the claim that the 30× parameter reduction and linear-time solver come with no quality loss.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two-stage phase reconstruction framework of [21] can be made drastically cheaper without a measured loss in quality. On the neural side, a new causal CNN with 8.46k parameters—roughly 30× smaller than the 247.81k-parameter predecessor—jointly predicts the FPD and BPD phase-derivative features at 0.27 GMAC/s, using a stem-body-head design with concatenated stem/body outputs, batch normalization, leaky ReLUs, and 1×1 convolutions. On the numerical side, the paper observes that the least-squares matrix A = Λ_{τ0} + D_{τ0}^H Γ_{τ0} D_{τ0} is tridiagonal and positive semidefinite when Λ and Γ are diagonal, so the system can be assembled directly from its diagonals and solved exactly with Thomas' algorithm in O(L) arithmetic and memory, orders of magnitude faster than direct inversion or LGMRES. An optional strided inference mode halves the CNN's cost by processing every other frame and emitting two outputs per pass, at the price of one hop of look-ahead. The experiments on LibriSpeech report ESTOI, WB-PESQ, and log-spectral convergence scores close to ground truth and statistically indistinguishable from the authors' reimplementation of [21], supporting the claim that the pipeline is online, efficient, and high-quality.

Load-bearing premise

The claim that the pipeline keeps full quality at much lower cost rests on the magnitude-weighting parameters inherited from the prior method being set correctly and on the authors' reimplementation of that prior method being faithful to the published version.

Editorial extensions

If this is right

  • At 0.27 GMAC/s and 8.46k parameters (0.14 GMAC/s with strided inference), the whole pipeline becomes small enough for low-power, on-device real-time use.
  • The strided inference mode gives system designers a direct trade-off: one hop of added latency halves the neural network's arithmetic cost, which is useful when compute or energy is scarcer than latency.
  • Replacing generic least-squares solves with the tridiagonal Thomas' algorithm removes the memory and time blowup of constructing and inverting a dense matrix, making the second stage exact and O(L) for any STFT resolution.
  • Across ESTOI, WB-PESQ, and log-spectral convergence, the authors measure no statistical quality drop relative to their implementation of the prior method, while staying close to ground-truth phase reconstruction.
  • Together the three changes realize a spectrogram inversion pipeline that is simultaneously online, computationally light, and high-quality, the paper's central claim.

Reading between the lines

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

  • Because the second-stage matrix is tridiagonal and positive semidefinite, the entire solve is differentiable and could be rolled into end-to-end training, letting the Λ and Γ weights be learned rather than inherited; the authors mention this only as future work.
  • The look-ahead striding scheme is a temporal subsampling trick that could be applied along the frequency axis as well, halving cost again at the price of frequency resolution—an unsupported extension the paper only lists as a possibility.
  • The tridiagonal structure is generic to any phase-derivative least-squares problem with the same bidiagonal D (i.e., any FPD-style feature), so the O(L) solver could be reused by other vocoders or phase postprocessors with different upstream networks.
  • Since the quality comparison is against the authors' own reimplementation of the prior model, the absolute scores may shift when the original implementation is used; the speedup itself is independent of that, but the 'no quality cost' wording is best read as relative to this specific baseline.
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 three efficiency improvements for online speech spectrogram inversion: a compact causal CNN with 8.46k parameters (about 30x smaller than the prior method), a strided inference scheme that halves CNN cost at the price of one hop of latency, and an O(L) tridiagonal solver for the second-stage least-squares phase estimation. The authors report that these innovations maintain the quality of the previous method [21] on Librispeech, as measured by ESTOI, WB-PESQ, and LSC, while drastically reducing computational cost.

Significance. If the claims hold, the paper makes a useful contribution to real-time spectrogram inversion by making the two-stage DNN-plus-least-squares approach of [21] much more computationally practical. The tridiagonal derivation in Eq. (12) is clean and correct, and the parameter reduction is substantial. However, the empirical support is weakened by missing definitions of the weighting matrices Λ and Γ, the use of a reimplementation rather than official code for the baseline, and the absence of statistical tests on the quality comparisons. The solver speedup is demonstrated on synthetic data, which is acceptable for a complexity claim but not fully transferable to real speech without a conditioning analysis.

major comments (4)
  1. [Section 3.3, Eqs. (10)-(11)] The weighting matrices Λτ0 and Γτ0 are stated to be 'diagonal nonnegative' and 'inherited from [21]', but their entries are never defined in this paper. These weights determine the phase-estimation objective, and different choices (e.g., identity versus magnitude-dependent weighting) will change the reconstructed phases and the reported ESTOI/PESQ/LSC scores. Because the evaluation in Figure 4a compares 'Proposed' to 'Prev.', the authors' own reimplementation of [21], the claim of 'no cost in output quality' is not fully reproducible and may not transfer to the actual published method if the reimplementation uses different weights. Please provide the exact definitions or reproduce the relevant equations from [21], and ideally release the code.
  2. [Section 3.3, Eq. (12) and following] The paper states that A is positive semidefinite and therefore Thomas' algorithm can be used. Positive semidefiniteness alone does not guarantee that all pivots are nonzero; Thomas' algorithm for tridiagonal systems requires positive definiteness or an equivalent pivot condition (e.g., strict diagonal dominance or irreducible diagonal dominance). If Λ can have zero diagonal entries, A may be singular, and the solver may divide by zero or produce unstable results. Please clarify conditions on Λ and Γ that ensure nonsingularity, or modify the solver to handle semidefinite systems (e.g., with a tiny regularization or fallback pivoting), and demonstrate that these conditions hold for the speech settings used in the experiments.
  3. [Section 4.3, Figure 4] The central quality claim—'comparable performance to [21]'—rests on boxplots without any statistical significance tests (e.g., paired Wilcoxon signed-rank tests) or confidence intervals. Moreover, the comparison is to the authors' own reimplementation of [21] ('Prev.'), not to an official implementation; if the reimplementation differs from the published method (particularly in the Λ/Γ settings), the comparison is not directly to the literature baseline. Please provide statistical tests over the 50 test utterances and clarify the fidelity of the reimplementation, including the exact Λ/Γ values used.
  4. [Section 4.2, Figure 3 (right)] The solver speedup is measured on synthetic systems with (u, v, Λ, Γ) sampled from i.i.d. standard Gaussian noise. This is a reasonable way to demonstrate complexity scaling, but the 'orders of magnitude' claim should be supported by measurements on systems derived from actual speech data, or at least by a report of the condition numbers of the synthetic matrices compared with those encountered in the speech evaluation. Otherwise, the reported speedup may be specific to the synthetic distribution rather than representative of real online use.
minor comments (5)
  1. [Abstract] The abstract contains a typo: 'we we observe' should read 'we observe'.
  2. [Section 3.1] The number '3 7.95 GMAC/s' appears to be a formatting error; it should likely read '37.95 GMAC/s' or similar.
  3. [Section 3.2] The strided inference scheme is described in text, but a small pseudocode or timing diagram would make the one-hop latency tradeoff and the exact mapping of skipped frames to outputs more explicit.
  4. [Section 4.2] The hardware description 'a commodity laptop' is vague; please specify the CPU model, memory, and operating system, and report the standard deviation of the runtime measurements in addition to the median and confidence interval.
  5. [Section 5] The limitations section is thorough regarding window/hop size and inconsistent spectrograms, but it does not mention that code is not released; given the reproducibility concerns with the Λ/Γ definitions, a note on code availability would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity is present: the O(L) solver claim follows from the algebra of Eq. 12, and the quality-maintenance claim is an empirical result on held-out Librispeech data.

full rationale

The efficiency claims are derived rather than assumed. Equation 12 expands D^H Γ D using the definition of D from Section 2.2 (a matrix with −u on the main diagonal and ones on the superdiagonal) and the diagonality of Λ and Γ, showing A = Λ + D^H Γ D has at most three nonzero diagonals; the O(L) complexity then follows from the standard tridiagonal-solver result of Golub and Van Loan [36]. This is a genuine first-principles structural derivation, not an input assumption, and the runtime benchmark in Figure 3 measures the proposed solver against direct inversion and LGMRES. The quality-maintenance claim, stated as "comparable performance to our own implementation of [21]" in Figure 4a, is empirical: both the proposed CNN and the reimplementation of [21]'s CNN are trained from scratch on Librispeech train-clean-360 and evaluated on the separate test-clean split with ESTOI, WB-PESQ, and LSC metrics, and no parameter is fitted to the test data, so no fitted input is renamed as a prediction. The framework (FPD/BPD features and the Eq. 10-11 objective) is inherited from [21] (Masuyama et al.), a different research group with no author overlap with this paper; the reference list contains no self-citations, so no load-bearing self-citation or imported-uniqueness pattern arises. The tridiagonality observation is derived in this paper and is not a renaming of a known result. Two legitimate concerns are reproducibility and numerical rigor rather than circularity: the weighting matrices Λ and Γ are never defined in this paper (only stated to be "diagonal nonnegative"), and positive-semidefiniteness alone does not guarantee that Thomas' algorithm is pivot-safe for singular systems; these belong to correctness risk and do not affect the circularity score.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central efficiency claims rest on mathematical structure (tridiagonality) plus standard algorithmic complexity. The main unstated inputs are the weighting matrices Λ and Γ from prior work, whose values are not given, and the assumption that the Gradient Theorem approximates phase derivatives under a Hann window. No new entities are introduced.

free parameters (1)
  • Λτ0 and Γτ0 diagonal weighting matrices = Not specified (inherited from [21])
    These nonnegative diagonal matrices control the least-squares objective in Eq. 10 and the closed-form solution in Eq. 11. Their values are never given in this paper, so output quality depends on undocumented choices from prior work.
assumptions (4)
  • domain assumption The Gradient Theorem relations (Eqs. 2-3) remain approximately valid for the Hann window used in the STFT.
    The theorem is derived for Gaussian windows; the paper uses a Hann window and relies on prior work [23] for the approximation. Invoked throughout Sections 2.2 and 3.
  • domain assumption All STFT coefficients Y[ω, τ] are nonzero, so phase ratios in Eqs. 8-9 are well-defined.
    The recursive phase estimation divides by magnitudes; zero coefficients would break the ratios. The paper notes 'assuming all Y[ω, τ] ≠ 0' in Section 2.2.
  • domain assumption The least-squares objective in Eq. 10 is a valid way to estimate phases from predicted derivatives.
    The weighted objective is imported from [21] and is not derived in this paper; its weighting matrices are unspecified.
  • domain assumption The tridiagonal matrix A is nonsingular (or safely solvable by Thomas' algorithm) for the encountered data.
    A is only guaranteed positive semidefinite, not positive definite; the paper does not discuss degenerate cases where a zero pivot could occur. Section 3.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Neural and Numerical Methods for High-Quality Online Speech Spectrogram Inversion via Gradient Theorem." pith.science (2026). https://pith.science/paper/MQWZWKWD

@misc{pith2026250524498,
  author       = {Pith},
  title        = {Pith review of: Efficient Neural and Numerical Methods for High-Quality Online Speech Spectrogram Inversion via Gradient Theorem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQWZWKWD}},
  note         = {Machine review of arXiv:2505.24498}
}
read the original abstract

Recent work in online speech spectrogram inversion effectively combines Deep Learning with the Gradient Theorem to predict phase derivatives directly from magnitudes. Then, phases are estimated from their derivatives via least squares, resulting in a high quality reconstruction. In this work, we introduce three innovations that drastically reduce computational cost, while maintaining high quality: Firstly, we introduce a novel neural network architecture with just 8k parameters, 30 times smaller than previous state of the art. Secondly, increasing latency by 1 hop size allows us to further halve the cost of the neural inference step. Thirdly, we we observe that the least squares problem features a tridiagonal matrix and propose a linear-complexity solver for the least squares step that leverages tridiagonality and positive-semidefiniteness, achieving a speedup of several orders of magnitude. We release samples online.

Figures

Figures reproduced from arXiv: 2505.24498 by the authors.

Figure 1
Figure 1. Overview of the proposed pipeline, modified from [21]. A causal, lightweight CNN is trained to jointly map STFT log-magnitudes into FPD and BPD features (defined in Section 2) at time frame τ0. Then, phases (Φ) are recursively estimated via complex least squares, computed in linear time w.r.t. number of frequency bins. The waveform can be then reconstructed via ISTFT. Additionally, the CNN can be modified to provide… view at source ↗
Figure 3
Figure 3. Summary of computational improvements (see Section 4.2 for extended discussion). Left: Our proposed CNN is ∼30× smaller and faster than the previous one from [21]. Right: Run￾times of different solvers for Equation 11, as a function of di￾mensionality L+1. Our proposed method is consistently fastest, and several orders of magnitude faster than other competitive matrix-free methods like LGMRES [37]. Lines show median… view at source ↗
Figure 2
Figure 2. Our proposed causal CNN for stage 1, featuring 8.46k parameters and 0.27 GMAC/s (Section 3.1). Tensor dimensions are Channels × Frequency bins × Time frames. The ⊙ nodes represent elementwise multiplication, and ⊕ con￾catenation across channel dimension. All convolutions are 2D, with unit stride. All padding is causal (left-sided) . Kernel sizes are detailed across (freq×time) dimensions. 3.2. Reducing CNN Computati… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance comparison for different approaches and metrics. See Section 4.3 for more details. in terms of quality and intelligibility, supporting the validity of our setup (VOCOS is worse here, but this may be due to the up￾sampling from 16kHz to 24kHz). Comparing Pre…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages

  1. [21]

    and with comparable performance

  2. [1]

    Introduction The Short-Term Fourier Transform (STFT) magnitudes of an audio waveform, also calledspectrograms, are a widely used representation in speech processing tasks such as recognition [1], denoising [2], separation [3], enhancement [4], and syn- thesis [5]. In this paper, we focus on the task of converting a spectrogram into a waveform by first est...

  3. [2]

    Then, phases (Φ) are recursively estimated via complex least squares, computed in linear time w.r.t

    at time frameτ 0. Then, phases (Φ) are recursively estimated via complex least squares, computed in linear time w.r.t. number of frequency bins. The waveform can be then reconstructed via ISTFT. Additionally, the CNN can be modified to provide also the features for frameτ -1 at minimal overhead, thus requiring only half the forward passes at the price of ...

  4. [3]

    A causal Convolutional Neural Network (CNN) architecture with only∼8k parameters,∼30×smaller and faster than

  5. [4]

    copy-synthesis

    Experiments and Discussion We ran several RSI pipelines to reconstruct clean speech wave- forms from their STFT log-magnitudes. We first show that our proposed changes to [21] result in a drastic reduction of compu- tation, both in terms of memory and runtime (Figure 3 and Sec- tion 4.2). We then show that our results are high-quality, with comparable per...

  6. [5]

    A modified inference scheme to optionally reduce computa- tion by an extra∼2×at the cost of one hop in latency

  7. [6]

    Efficient Neural and Numerical Methods for High-Quality Online Speech Spectrogram Inversion via Gradient Theorem

    A characterization of the second stage as a tridiagonal and positive-semidefinite linear system, leading to orders of mag- nitude faster computation with guaranteed linear complexity. These innovations allow us to achieve a speech spectrogram inversion solution that isonline, efficient and high-quality. Sec- tion 2 provides further context. We describe ou...

  8. [7]

    Online Spectrogram Inversion via Gradient Theorem Consider the STFT of a functiony(t)∈L 2(R)w.r.t

    Background 2.1. Online Spectrogram Inversion via Gradient Theorem Consider the STFT of a functiony(t)∈L 2(R)w.r.t. window h(t)∈L2(R), both real-valued: Yy,h(ω, t):= Z t′∈R y(t+t ′)h(t ′)e −2πiωt dt′, ω, t∈R(1) If we take a Gaussian window in the formφ λ(t):=e−π t2 λ , the Gradient Theorem (GT) expresses the relation between log- magnitude and phase ofYas ...

Show all 57 references
  1. [8]

    Efficient Neural and Numerical Methods We now describe our innovations to [21], aimed at reducing computation without affecting performance. 3.1. An Efficient CNN for the First Stage The CNNs introduced in [21] comprise 7 learnable layers, fea- turing a series of residual, sig...

  2. [9]

    Signal reconstruction from stft mag- nitude : A state of the art,

    N. Sturmel and L. Daudet, “Signal reconstruction from stft mag- nitude : A state of the art,” inDAFx, 2011

  3. [10]

    On phase retrieval of finite-length sequences using the initial time sample,

    H. Sahinoglou and S. Cabrera, “On phase retrieval of finite-length sequences using the initial time sample,”IEEE Transactions on Circuits and Systems, vol. 38, no. 8, pp. 954–958, 1991

  4. [11]

    While results are good for our settings, other window and hop sizes, as well as enhanced/inconsistent spectrograms, remain to be evaluated

    Limitations and Future Work In this work we introduced three innovations on top of the framework from [21] in order to achieve drastic reduction in computation, while maintaining low latency and high quality. While results are good for our settings, other window and hop sizes,...

  5. [12]

    This work was done while AF was a re- search scientist intern at Meta Reality Labs

    Acknowledgements The authors thank Buye Xu, Sanjeel Parekh, Michael Mc- Manus, Adrian Stepien and Kuba Rad for the constructive and helpful discussions. This work was done while AF was a re- search scientist intern at Meta Reality Labs

  6. [13]

    Deep speech 2: end-to-end speech recognition in English and Mandarin,

    D. Amodeiet al., “Deep speech 2: end-to-end speech recognition in English and Mandarin,” inICML, 2016

  7. [14]

    Denoising speech based on deep learning and wavelet decomposition,

    L. Wanget al., “Denoising speech based on deep learning and wavelet decomposition,”Scientific Programming, vol. 2021, no. 1, p. 8677043, 2021

  8. [15]

    End-to-end speech separation with unfolded iterative phase reconstruction,

    Z.-Q. Wanget al., “End-to-end speech separation with unfolded iterative phase reconstruction,” inInterspeech 2018, 2018

  9. [16]

    T-GSA: Transformer with Gaussian-weighted self-attention for speech enhancement,

    J. Kim, M. El-Khamy, and J. Lee, “T-GSA: Transformer with Gaussian-weighted self-attention for speech enhancement,” in ICASSP, 2020

  10. [17]

    GANSynth: Adversarial neural audio synthesis,

    J. Engelet al., “GANSynth: Adversarial neural audio synthesis,” inICLR, 2019

  11. [18]

    The importance of phase in speech enhancement,

    K. Paliwal, K. W ´ojcicki, and B. Shannon, “The importance of phase in speech enhancement,”Speech Communication, vol. 53, no. 4, pp. 465–494, 2011

  12. [19]

    Phasebook and friends: Leveraging discrete representations for source separation,

    J. Le Rouxet al., “Phasebook and friends: Leveraging discrete representations for source separation,”IEEE Journal of Selected Topics in Signal Processing, vol. 13, no. 2, pp. 370–382, 2019

  13. [20]

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

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

  14. [22]

    Magnitude-phase relationships for short-time Fourier transforms based on Gaussian analysis windows,

    M. Portnoff, “Magnitude-phase relationships for short-time Fourier transforms based on Gaussian analysis windows,” in ICASSP, 1979

  15. [23]

    Signal reconstruction from short-time Fourier transform magnitude,

    S. Nawab, T. Quatieri, and J. Lim, “Signal reconstruction from short-time Fourier transform magnitude,”IEEE Transactions on Acoustics, Speech, and Sig. Proc., vol. 31, no. 4, 1983

  16. [24]

    PhaseLift: Exact and stable signal recovery from magnitude measurements via con- vex programming,

    E. J. Cand `es, T. Strohmer, and V . V oroninski, “PhaseLift: Exact and stable signal recovery from magnitude measurements via con- vex programming,”Communications on Pure and Applied Math- ematics, vol. 66, no. 8, pp. 1241–1274, 2013

  17. [25]

    Estimating a signal from a magni- tude spectrogram via convex optimization,

    D. L. Sun and J. O. S. III, “Estimating a signal from a magni- tude spectrogram via convex optimization,”Journal of the Audio Engineering Society, Oct 2012

  18. [26]

    An efficient algorithm for real-time spectrogram inversion,

    G. T. Beauregard, X. Zhu, and L. Wyse, “An efficient algorithm for real-time spectrogram inversion,” inDAFx, 2005

  19. [27]

    Signal estimation from modified short- time Fourier transform,

    D. Griffin and J. Lim, “Signal estimation from modified short- time Fourier transform,” inICASSP, 1983

  20. [28]

    A flexible online framework for projection-based STFT phase retrieval,

    T. Peeret al., “A flexible online framework for projection-based STFT phase retrieval,” inICASSP, 2024

  21. [29]

    Inversion of auditory spectrograms, tradi- tional spectrograms, and other envelope representations,

    R. Decorsi `ereet al., “Inversion of auditory spectrograms, tradi- tional spectrograms, and other envelope representations,”TASLP, vol. 23, no. 1, pp. 46–56, 2015

  22. [30]

    Phase retrieval with Bregman divergences and application to audio signal recovery,

    P.-H. Vialet al., “Phase retrieval with Bregman divergences and application to audio signal recovery,”IEEE Journal of Selected Topics in Signal Processing, vol. 15, no. 1, pp. 51–64, 2021

  23. [31]

    Single pass spec- trogram inversion,

    G. T. Beauregard, M. Harish, and L. Wyse, “Single pass spec- trogram inversion,” inIEEE International Conference on Digital Signal Processing (DSP), 2015

  24. [32]

    Phase-locked vocoder,

    M. Puckette, “Phase-locked vocoder,” inWASPAA, 1995, pp. 222– 225

  25. [33]

    Online phase reconstruction via DNN-based phase differences estimation,

    Y . Masuyamaet al., “Online phase reconstruction via DNN-based phase differences estimation,”TASLP, vol. 31, pp. 163–176, 2023

  26. [34]

    Empirical evaluation of rectified activations in con- volutional network,

    B. Xuet al., “Empirical evaluation of rectified activations in con- volutional network,” inarXiv, no. 1505.00853, 2015

  27. [35]

    A noniterative method for reconstruction of phase from STFT magnitude,

    Z. Pr ˚uˇsa, P. Bal´azs, and P. L. Søndergaard, “A noniterative method for reconstruction of phase from STFT magnitude,”TASLP, vol. 25, no. 5, pp. 1154–1164, 2017

  28. [36]

    Real-time spectrogram inversion using phase gradient heap integration,

    Z. Pr ˚uˇsa and P. L. Søndergaard, “Real-time spectrogram inversion using phase gradient heap integration,” inDAFx, 2016

  29. [37]

    WaveNet: A generative model for raw audio,

    A. van den Oordet al., “WaveNet: A generative model for raw audio,” in9th ISCA Speech Synthesis Workshop, 2016

  30. [38]

    HiFi-GAN: generative adversar- ial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: generative adversar- ial networks for efficient and high fidelity speech synthesis,” in NeurIPS, 2020

  31. [39]

    VOCOS: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis,

    H. Siuzdak, “VOCOS: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis,” inICLR, 2024

  32. [40]

    Phase reconstruction from amplitude spec- trograms based on V on-Mises-distribution deep neural network,

    S. Takamichiet al., “Phase reconstruction from amplitude spec- trograms based on V on-Mises-distribution deep neural network,” IWAENC, 2018

  33. [41]

    Phase reconstruction from amplitude spectrograms based on directional-statistics deep neural networks,

    ——, “Phase reconstruction from amplitude spectrograms based on directional-statistics deep neural networks,”Elsevier Signal Processing, vol. 169, no. C, Apr. 2020

  34. [42]

    Recurrent phase recon- struction using estimated phase derivatives from deep neural net- works,

    L. Thieling, D. Wilhelm, and P. Jax, “Recurrent phase recon- struction using estimated phase derivatives from deep neural net- works,” inICASSP, 2021

  35. [43]

    to measure speech intelligibility (0-100, higher is better), WB-PESQ [44] for speech quality (1.04-4.64, higher is better10) and Log-Spectral Convergence (LSC) [21] for Euclidean error (dB, lower is better). In Figure 4a, we observe that our custom implementations (Proposedand...

  36. [44]

    Inter-frequency phase difference for phase reconstruction using deep neural networks and maximum likeli- hood,

    N. B. Thienet al., “Inter-frequency phase difference for phase reconstruction using deep neural networks and maximum likeli- hood,”TASLP, vol. 31, 2023

  37. [45]

    Two-stage phase reconstruction using DNN and von Mises distribution-based maximum likelihood,

    ——, “Two-stage phase reconstruction using DNN and von Mises distribution-based maximum likelihood,” inAPSIPA ASC, 2021

  38. [46]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inICML, 2015

  39. [47]

    J. W. Demmel,Applied Numerical Linear Algebra. Society for Industrial and Applied Mathematics, 1997

  40. [48]

    G. H. Golub and C. F. Van Loan,Matrix Computations. The Johns Hopkins University Press, 2013

  41. [49]

    A technique for accelerating the convergence of restarted GMRES,

    A. H. Baker, E. R. Jessup, and T. Manteuffel, “A technique for accelerating the convergence of restarted GMRES,”SIAM Journal on Matrix Analysis and Applications, vol. 26, no. 4, 2005

  42. [50]

    Librispeech: An ASR corpus based on public domain audio books,

    V . Panayotovet al., “Librispeech: An ASR corpus based on public domain audio books,” inICASSP, 2015

  43. [51]

    Delving deep into rectifiers: Surpassing human- level performance on imagenet classification,

    K. Heet al., “Delving deep into rectifiers: Surpassing human- level performance on imagenet classification,” inICCV, 2015

  44. [52]

    On the variance of the adaptive learning rate and beyond,

    L. Liuet al., “On the variance of the adaptive learning rate and beyond,” inICLR, 2020

  45. [53]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” inICLR, 2017

  46. [54]

    Subspectral normalization for neural audio data processing,

    S. Changet al., “Subspectral normalization for neural audio data processing,”ICASSP, 2021

  47. [55]

    An algorithm for predicting the intelli- gibility of speech masked by modulated noise maskers,

    J. Jensen and C. H. Taal, “An algorithm for predicting the intelli- gibility of speech masked by modulated noise maskers,”TASLP, vol. 24, no. 11, pp. 2009–2022, 2016. [44]Rec. P.862.2: Wideband extension to Recommendation P.862 for the assessment of wideband telephone networks...

  48. [56]

    Phase reconstruction in single chan- nel speech enhancement based on phase gradients and estimated clean-speech amplitudes,

    Y . Song and N. Madhu, “Phase reconstruction in single chan- nel speech enhancement based on phase gradients and estimated clean-speech amplitudes,” inICASSP, 2024

  49. [57]

    Fast spectrogram inversion using multi-head convolutional neural networks,

    S. ¨O. Arık, H. Jun, and G. Diamos, “Fast spectrogram inversion using multi-head convolutional neural networks,”IEEE Signal Processing Letters, vol. 26, no. 1, pp. 94–98, 2019

Pith tools

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