Pith. sign in

REVIEW 3 major objections 4 minor 49 references

Learnable Adaptive Time-Frequency Representation via Differentiable Short-Time Fourier Transform

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that the short-time Fourier transform can be made differentiable with respect to real-valued window and hop lengths, enabling gradient-based and end-to-end learning of spectrogram parameters.

desk verdict A mostly sound and honest consolidation of the differentiable STFT program, with one concrete off-by-one error in the hop-length gradient formula that needs fixing before the claims hold as written. read the letter →

arxiv 2506.21440 v1 pith:LQV3Y4CB submitted 2025-06-26 cs.SD cs.LGeess.ASeess.SP

Maxime Leiber , Yosra Marnissi , Axel Barrau , Sylvain Meignen , Laurent Massoulié This is my paper · ORCID
classification cs.SDcs.LGeess.ASeess.SP
keywords short-timeFouriertransformspectrogramdifferentiableSTFTwindowlengthoptimizationhopgradient-basedlearningtime-frequencyrepresentationbackpropagation
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

The paper claims that the standard spectrogram pipeline can be made fully differentiable: by treating the analysis window length and the hop length as real-valued parameters, every stage of the short-time Fourier transform admits analytic gradients. This turns the choice of window and hop sizes from a discrete search over candidate values into a continuous optimization problem that can be solved by gradient descent. The paper also lets the window length vary separately for each time-frequency cell and lets frame positions move continuously, which gives the representation far more flexibility than a fixed-grid spectrogram. A sympathetic reader would care because this is the missing link for tuning time-frequency representations automatically, and for training them jointly with neural networks on a downstream task.

What carries the argument

The mechanism is the contracted-window family $\omega(x,\theta)$ together with the generalized STFT operator $\mathcal{S}: \mathbb{R}^N \times ]0,L]^{M\times N} \to \mathbb{C}^{M\times N}$ that maps a set of real-valued temporal positions and per-time-frequency window lengths to an $M\times N$ complex matrix. The key identity is that the partial derivative of the transform with respect to any parameter is another DSTFT evaluated with the derivative of the window, so all backpropagation formulas reduce to matrix multiplications with the same dimension as the forward pass. This is what makes the parameter learning tractable and exact.

What would settle it

Run the proposed optimization on a multicomponent signal from many random initializations of the time-frequency window-length map; if the runs converge to materially different spectrograms, the entropy landscape is non-convex and the claimed optimum is not unique.

Watch

Extended reading notes

Core claim

A key contribution of this work is the demonstration that the STFT is differentiable with respect to its window and hop lengths, provided these parameters are treated as real-valued. The paper constructs a family of contracted windows $\omega(x,\theta) = (L/\theta)\omega_L(Lx/\theta)$ and a generalized STFT in which the window length $\theta_{m,n}$ can vary with both time and frequency while the frame positions $t_n$ are real-valued. For this transform the partial derivatives with respect to every parameter are themselves STFTs computed with differentiated window functions, so the backward pass has the same cost and structure as the forward pass. Consequently any almost-everywhere-smooth loss on the spectrogram can be minimized by gradient descent, and the same chain rule lets the STFT act as a trainable first layer of a neural network.

Load-bearing premise

The paper's optimality claims rest on the unproven assumption that the Shannon-entropy loss is convex in the window length, so that gradient descent cannot settle in local optima.

Editorial extensions

If this is right

  • Spectrogram-based systems can replace grid search over window and hop lengths with gradient descent, removing the exponential cost of optimizing time-frequency-varying parameters.
  • The STFT becomes a differentiable layer, so its parameters can be trained jointly with a neural network on the task loss rather than tuned in advance.
  • Window lengths are no longer confined to a discrete candidate set, so the optimized representation can in principle be more accurate than any discrete choice.
  • Representation objectives such as Shannon entropy and task objectives such as frequency-tracking error are minimized in the same framework with the same backward pass.
  • The backward pass matches the forward pass in complexity, making each parameter-learning step roughly twice the cost of a single STFT computation.

Reading between the lines

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

  • If the differentiability result holds, it likely extends beyond the STFT: the same contraction trick could be applied to other parameterized transforms such as the S-transform or wavelet scalograms, turning their scale parameters into trainable variables.
  • The per-time-frequency window-length map learned by the DSTFT could be interpreted as a data-dependent resolution mask, which might serve as a regularizing prior for non-stationary signal analysis.
  • A direct test of the approach's benefit would be to compare one network trained with a learnable window against a standard fixed-window pipeline on a larger audio benchmark; the paper's small spoken-digit experiment suggests but does not establish a general advantage.
  • The analytic backpropagation formulas could serve as a cheap oracle for verifying automatic-differentiation implementations of spectrogram layers in deep-learning frameworks.
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

3 major / 4 minor

Summary. The paper proposes a differentiable reformulation of the short-time Fourier transform (DSTFT) in which the window length θ_{m,n} and the temporal positions t_n (equivalently the hop lengths H_n = t_n − t_{n−1}) are treated as real-valued trainable parameters. A contracted window family is introduced in Section III, and analytic partial derivatives of the STFT with respect to window length and temporal position/hop length are derived in Section IV, together with backpropagation formulas and a complexity analysis in Section V. The framework is demonstrated in two representation-driven tasks (time-frequency varying window length; time-varying window and hop lengths) and two task-driven tasks (frequency tracking; joint optimization with a CNN for spoken digit classification on FSDD). Code is provided on GitHub.

Significance. The central idea is useful: if the STFT is made differentiable with respect to its analysis parameters, then spectrogram-based pipelines can tune window and hop lengths by gradient descent, including joint training with neural networks. A genuine strength is that the derivative formulas in Section IV express every gradient as a DSTFT with a modified window, so the backward pass has the same complexity as the forward pass and can be implemented exactly rather than through generic autodiff. The paper also ships reproducible code and reports a small but sensible FSDD experiment in which a jointly optimized window length improves test accuracy over all fixed window lengths (80.7% vs 79.7%). However, the time-varying hop-length gradient in Eq. (15) has an off-by-one error, and several optimality claims in Section VI rest on an unproved convexity assertion. These issues are localizable and fixable, but they affect load-bearing parts of the paper as written.

major comments (3)
  1. [Section IV.B, Eq. (15)] Equation (15) is not correct as written for the time-varying hop length. Since t_{n'} = Σ_{i=0}^{n'} H_i, one has ∂t_{n'}/∂H_n = 1 for every n' ≥ n, including n' = n. Therefore Eq. (14) implies ∂_{H_n} S(Ω) = (−S_{∂xω}(t_{n'}, m', θ_{m',n'}) · 1_{n'≥n})_{m',n'}. The published expression uses 1_{x>0}(n'−n), i.e., n' > n, which drops the self-frame term n' = n. As a result, any loss that depends on S(t_n) receives a biased gradient for H_n. The constant-hop formulas in Eqs. (16) and (24) are consistent, and the experiments in Section VI.B optimize t_n rather than H_n directly, so the empirical demonstrations may survive; nevertheless the paper's advertised differentiability with respect to the hop length is not correctly established. This should be fixed by replacing the indicator with 1_{n'≥n} and updating the surrounding text and backpropagation relation in Eq. (20).
  2. [Section VI.A, Eq. (26) and Fig. 2] The paper states that numerical analysis 'suggests that the loss function is convex with respect to θ, ensuring the uniqueness of the optimum,' and explicitly defers a proof to future work. A single plotted loss curve is not a proof of convexity, and the assertion is load-bearing: it is used to label the constant-window result 'optimal,' to justify initializing the TF-varying optimization with the constant-window optimum, and to argue that λ → +∞ regularization recovers a convex problem. Unless a proof or a precise local-convexity statement is supplied, the optimality claims in this section should be reworded to describe the results as local optima of the chosen criterion.
  3. [Section VI.B, Eq. (32)] The claim that C(Ω) in Eq. (32) is 'differentiable with respect to Ω, anywhere in the parameter space' is incorrect. The term min(t_{n+1} − t_n, (θ_{n+1} + θ_n)/2) is nonsmooth at the equality boundary, and the two indicator functions 1_{x<L_s}(·) and 1_{x>0}(·) are discontinuous step functions. Since the experiment in Section VI.B minimizes the loss in Eq. (29) that includes C(Ω), this is not purely cosmetic. The authors should either replace C by a smooth approximation, explicitly state that a subgradient method is used, or restrict the differentiability claim to the relevant open set where the indicators are constant and the min branch is unique.
minor comments (4)
  1. [Section IV.A, Eqs. (11) and (12)] The labels in Eqs. (11) and (12) appear interchanged: Eq. (11) uses θ_n and δ_{n,n'} and is therefore the time-varying case, while Eq. (12) uses θ_m and δ_{m,m'} and is the frequency-varying case. The text also swaps the domain dimensions. The backpropagation formulas in Section IV.C use the correct labeling and should be harmonized with these equations.
  2. [Section V.A, after Eq. (25)] The statement that 'the tapering function ω is zero at the boundaries of its support' is true for the Hann window in Eq. (4) but false for the truncated Gaussian window in Eq. (5), where the window value at x = ±θ/2 is nonzero. The summation limits and the boundary discussion should be qualified to the window family actually used, or the truncated Gaussian should be smoothed.
  3. [Table III] The column headers in Table III repeat 'train loss, val loss, test loss' from Table II, but the entries are classification accuracies in percent. The headers should read 'train accuracy, val accuracy, test accuracy.'
  4. [Section III.A, Eq. (5)] The paper acknowledges that differentiability at the exact boundaries of the compact support is a theoretical point not addressed further. This is acceptable for the Hann experiments, but the truncated Gaussian example in Eq. (5) does not satisfy the blanket differentiability assumption stated after Eq. (6); the assumption should be stated as a hypothesis of the derivation rather than as a property of the examples.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the derivative derivation is self-contained, with only minor self-citation and a self-referential MSE demo.

full rationale

The core derivation is self-contained. Section III defines a contracted window family and a generalized STFT with real-valued temporal positions and window lengths, and Section IV computes the derivatives S_{∂θω} and S_{∂xω} via the chain rule under the stated differentiability assumption. None of the formulas (9)-(16) or backpropagation identities (18)-(24) is fitted from data or imported from the authors' prior work; [15]-[17] are cited as historical precursors, and the paper re-derives the necessary expressions. The CNN experiment in Sec VII.B is evaluated on a held-out test set (80.7% vs 79.7% for the best fixed window), so that claim does not reduce to its training objective. The task-driven frequency-tracking demo in Sec VII.A is weaker: it optimizes the same MSE loss that it displays, so the 'effectiveness' shown there is close to an optimization demonstration rather than an independent prediction; however, the paper does not present a quantitative held-out improvement in that section, so this is an evaluation weakness rather than a circular derivation. Two non-circular caveats should be noted: the Sec VI.A convexity claim is unproven and explicitly deferred to future work, and Eq. (15) appears to use the indicator n'>n where the chain rule requires n'≥n, which is a correctness issue for the hop-length gradient as written, not a circularity. Overall, the central mathematical claim has independent content and is not forced by definition or by self-citation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central derivation is nearly self-contained: it relies on standard Fourier analysis plus three domain assumptions (normalized window family, differentiability of the window, almost-everywhere smooth losses). The only assumption ad hoc to this paper is the asserted convexity of the Shannon entropy in θ, which is unproven and load-bearing for the optimality claims. No hidden physical or model entities are introduced; the DSTFT is an explicit continuous re-parameterization of the classical STFT (Eqs. 2-7), and its per-time-frequency window field θ_{m,n} is a construction with falsifiable handles (held-out classification and tracking experiments, public repository), not an unexplained postulate. No free parameters are hidden in the derivation itself; the fitted values are the learned STFT parameters, the manually chosen λ, and untabulated optimizer settings.

free parameters (3)
  • learned window length θ = θ=34.9 samples on FSDD; per-bin and per-time fields in Secs. VI.A-VI.C (values shown as images only)
    The paper's optimization targets. θ=34.9 is the jointly trained window length in Sec. VII.B. These are fitted to data by construction, so any downstream claim depends on them.
  • regularization weight λ = 1e-3
    Hand-chosen in Sec. VI.A (Eq. 27) to balance Shannon entropy against the total-variation penalty; the paper gives no sensitivity analysis.
  • optimization hyperparameters = not tabulated
    Learning rates, iteration counts P (10-100 constant, 50-300 TF-varying), overlap factor α, and the 20 dB noise level for the vibration signal are set by hand; the paper states P depends on these choices.
assumptions (5)
  • standard math Standard discrete Fourier transform theory and FFT complexity
    Invoked throughout Sec. II and in the complexity analysis of Sec. V.B.
  • domain assumption The normalization L/θ in the contracted window family keeps the DC Fourier gain constant across θ
    Eq. (2) and the pure-tone argument in Sec. III.A. This normalization defines fair comparison between STFTs of different window lengths; a different normalization would change the entropy loss landscape.
  • domain assumption The window function ω(x,θ) is differentiable in (x,θ) on its support, boundary behavior excluded
    Stated in Sec. III.A ('we assume this property holds'); for the truncated Gaussian of Eq. (5) the support boundary has a jump of size exp(-9π), which the paper explicitly sets out of scope.
  • ad hoc to paper The Shannon entropy loss is convex in θ for a constant window
    Sec. VI.A: inferred from Fig. 2 ('numerical analysis'), proof deferred to future work; used to justify uniqueness of the optimum and the initialization strategy for the TF-varying optimization.
  • domain assumption Loss functions are almost everywhere smooth, so gradients exist
    Sec. IV.C restricts L to 'almost everywhere smooth differentiable' losses; the coverage penalty C(Ω) in Eq. (32) is then claimed differentiable anywhere, which is stronger than the min and indicator construction supports.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learnable Adaptive Time-Frequency Representation via Differentiable Short-Time Fourier Transform." pith.science (2026). https://pith.science/paper/LQV3Y4CB

@misc{pith2026250621440,
  author       = {Pith},
  title        = {Pith review of: Learnable Adaptive Time-Frequency Representation via Differentiable Short-Time Fourier Transform},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQV3Y4CB}},
  note         = {Machine review of arXiv:2506.21440}
}
read the original abstract

The short-time Fourier transform (STFT) is widely used for analyzing non-stationary signals. However, its performance is highly sensitive to its parameters, and manual or heuristic tuning often yields suboptimal results. To overcome this limitation, we propose a unified differentiable formulation of the STFT that enables gradient-based optimization of its parameters. This approach addresses the limitations of traditional STFT parameter tuning methods, which often rely on computationally intensive discrete searches. It enables fine-tuning of the time-frequency representation (TFR) based on any desired criterion. Moreover, our approach integrates seamlessly with neural networks, allowing joint optimization of the STFT parameters and network weights. The efficacy of the proposed differentiable STFT in enhancing TFRs and improving performance in downstream tasks is demonstrated through experiments on both simulated and real-world data.

Figures

Figures reproduced from arXiv: 2506.21440 by the authors.

Figure 3
Figure 3. Top row: Spectrogram (left) and its zoomed-in section (right) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 1
Figure 1. Spectrograms (left) and zoomed-in spectrograms (right) with respec [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Loss function corresponding to Eq. (26) with respect to window length [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: In the first row, frequencies (left) and temporal signal (right). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Single window (left) and time-varying window (right) Spectrograms [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 5
Figure 5. Figure 5: The position of the tapering windows can smoothly shift along the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Loss per window length (top), angular speed associated with optimal [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 46 canonical work pages

  1. [16]

    Differentiable adaptive short-time Fourier transform with respect to the window length,

    M. Leiber, Y . Marnissi, A. Barrau, and M. El Badaoui, “Differentiable adaptive short-time Fourier transform with respect to the window length,” in IEEE int. conf. on Acoustics, Speech and Signal Process. (ICASSP), 2023

    work page 2023

  2. [17]

    Differentiable short-time Fourier transform with respect to the hop length,

    ——, “Differentiable short-time Fourier transform with respect to the hop length,” in IEEE Workshop on Statistical Signal Process. (SSP) , 2023

    work page 2023

  3. [1]

    Long-range acoustic detection and localization of blue whale calls in the northeast pacific ocean,

    K. M. Stafford, C. G. Fox, and D. S. Clark, “Long-range acoustic detection and localization of blue whale calls in the northeast pacific ocean,” The Journal of the Acoustical Society of America , vol. 104, no. 6, pp. 3616–3625, 1998

    work page 1998

  4. [2]

    ECG arrhythmia classification using STFT-based spectrogram and convolutional neural network,

    J. Huang, B. Chen, B. Yao, and W. He, “ECG arrhythmia classification using STFT-based spectrogram and convolutional neural network,”IEEE access, vol. 7, pp. 92 871–92 880, 2019

    work page 2019

  5. [3]

    A multi-order probabilistic approach for instantaneous angular speed tracking debriefing of the CMMNO14 diagnosis contest,

    Q. Lecl `ere, H. Andr ´e, and J. Antoni, “A multi-order probabilistic approach for instantaneous angular speed tracking debriefing of the CMMNO14 diagnosis contest,” Mechanical Systems and Signal Pro- cess., vol. 81, pp. 375–386, 2016

    work page 2016

  6. [4]

    Ast: Audio spectrogram trans- former,

    Y . Gong, Y .-A. Chung, and J. Glass, “Ast: Audio spectrogram trans- former,” in Interspeech 2021, 2021, pp. 571–575

    work page 2021

  7. [5]

    Unsupervised speech en- hancement with diffusion-based generative models,

    B. Nortier, M. Sadeghi, and R. Serizel, “Unsupervised speech en- hancement with diffusion-based generative models,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 481–12 485

    work page 2024

  8. [6]

    Improved musical onset detection with convolutional neural networks,

    J. Schl ¨uter and S. B ¨ock, “Improved musical onset detection with convolutional neural networks,” in 2014 IEEE Int. Conf. Acoust., speech and Signal Process. (ICASSP) . IEEE, 2014, pp. 6979–6983

    work page 2014

Show all 49 references
  1. [7]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” in Proc. Interspeech 2019 , 2019, pp. 2613–2617

    2019

  2. [8]

    Hybrid spectrogram and waveform source separation,

    A. D ´efossez, “Hybrid spectrogram and waveform source separation,” arXiv preprint arXiv:2111.03600 , 2021

    2021 arXiv

  3. [9]

    Introduction to digital speech processing,

    L. R. Rabiner, R. W. Schafer et al. , “Introduction to digital speech processing,” Foundations and Trends® in Signal Process. , vol. 1, no. 1–2, pp. 1–194, 2007

    2007

  4. [10]

    The role of window length and shift in complex-domain DNN-based speech enhancement,

    C. Garc ´ıa-Ruiz, A. M. Gomez, and J. M. Mart ´ın-Do˜nas, “The role of window length and shift in complex-domain DNN-based speech enhancement,” in IberSPEECH 2022, 2022, pp. 146–150

    2022

  5. [11]

    An empirical study on analysis window functions for text-independent speaker recognition,

    B. Barai, N. Das, S. Basu, and M. Nasipuri, “An empirical study on analysis window functions for text-independent speaker recognition,” int. Journal of Speech Technology , vol. 26, no. 1, pp. 211–220, 2023

    2023

  6. [12]

    Deep learning: Methods and applications,

    L. Deng and D. Yu, “Deep learning: Methods and applications,” Founda- tions and Trends® Signal Process. , vol. 7, no. 3-4, pp. 197–387, 2014

    2014

  7. [13]

    K. M. Prabhu, Window functions and their applications in Signal Process. Taylor & Francis, 2014. 12

    2014

  8. [14]

    Havin and B

    V . Havin and B. J¨oricke, The uncertainty principle in harmonic analysis. Springer Science & Business Media, 2012, vol. 28

    2012

  9. [15]

    A differentiable short-time Fourier transform with respect to the window length,

    M. Leiber, A. Barrau, Y . Marnissi, and D. Abboud, “A differentiable short-time Fourier transform with respect to the window length,” in European Signal Process. conf. (EUSIPCO) , 2022, pp. 1392–1396

    2022

  10. [18]

    The backpropagation algorithm,

    R. Rojas and R. Rojas, “The backpropagation algorithm,” Neural net- works: a systematic introduction , pp. 149–182, 1996

    1996

  11. [19]

    Backpropagation neural networks: a tutorial,

    B. J. Wythoff, “Backpropagation neural networks: a tutorial,” Chemo- metrics and Intelligent Laboratory Systems , vol. 18, no. 2, pp. 115–155, 1993

    1993

  12. [20]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015

    2015

  13. [21]

    Adaptive short-time Fourier analy- sis,

    R. N. Czerwinski and D. L. Jones, “Adaptive short-time Fourier analy- sis,” in IEEE Signal Process. Letters , vol. 4, 1997, pp. 42–45

    1997

  14. [22]

    Variable short-time Fourier transform for vibration signals with transients,

    J.-Y . Lee, “Variable short-time Fourier transform for vibration signals with transients,” Journal of Vibration and Control , vol. 21, no. 7, pp. 1383–1397, 2015

    2015

  15. [23]

    Improved instantaneous frequency estimation using an adaptive short-time Fourier transform,

    H. K. Kwok and D. L. Jones, “Improved instantaneous frequency estimation using an adaptive short-time Fourier transform,” IEEE Trans. on Signal Process. , vol. 48, pp. 2964–2972, 2000

    2000

  16. [24]

    Time-frequency representation based on an adaptive short-time Fourier transform,

    J. Zhong and Y . Huang, “Time-frequency representation based on an adaptive short-time Fourier transform,” in IEEE Trans. on Signal Process., vol. 58, 2010, pp. 5118–5128

    2010

  17. [25]

    STFT with adaptive window width based on the chirp rate,

    S. Pei and S. Huang, “STFT with adaptive window width based on the chirp rate,” IEEE Trans. on Signal Process. , vol. 60, no. 8, pp. 4065– 4080, 2012

    2012

  18. [26]

    The- ory, implementation and applications of nonstationary Gabor frames,

    P. Balazs, M. D ¨orfler, F. Jaillet, N. Holighaus, and G. Velasco, “The- ory, implementation and applications of nonstationary Gabor frames,” Journal of computational and applied mathematics , vol. 236, no. 6, pp. 1481–1496, 2011

    2011

  19. [27]

    Optimizing short-time Fourier transform parameters via gradient descent,

    A. Zhao, K. Subramani, and P. Smaragdis, “Optimizing short-time Fourier transform parameters via gradient descent,” in IEEE int. conf. on Acoustics, Speech and Signal Process. (ICASSP), 2021, pp. 736–740

    2021

  20. [28]

    Differentiable short-time Fourier transform window length selection driven by cyclo-stationarity,

    D. G. Marx and K. Gryllias, “Differentiable short-time Fourier transform window length selection driven by cyclo-stationarity,” in Proceedings of the Annual conf. of the Prognostics and Health Management Society, PHM, vol. 15. PHM Society, 2023, pp. 1–10

    2023

  21. [29]

    Idsn: A one-stage interpretable and dif- ferentiable stft domain adaptation network for traction motor of high- speed trains cross-machine diagnosis,

    C. He, H. Shi, and J. Li, “Idsn: A one-stage interpretable and dif- ferentiable stft domain adaptation network for traction motor of high- speed trains cross-machine diagnosis,” Mechanical Systems and Signal Process., vol. 205, p. 110846, 2023

    2023

  22. [30]

    A new method for the numerical analysis of non-stationary signals,

    K. Kodera, C. De Villedary, and R. Gendrin, “A new method for the numerical analysis of non-stationary signals,” Physics of the Earth and Planetary Interiors, vol. 12, no. 2-3, pp. 142–150, 1976

    1976

  23. [31]

    Analysis of time-varying signals with small BT values,

    K. Kodera, R. Gendrin, and C. Villedary, “Analysis of time-varying signals with small BT values,” IEEE Trans. on Acoustics, Speech, and Signal Process., vol. 26, no. 1, pp. 64–76, 1978

    1978

  24. [33]

    Time frequency reas- signment: from principles to algorithms,

    P. Flandrin, F. Auger, and E. Chassande-Mottin, “Time frequency reas- signment: from principles to algorithms,” Applications in time-frequency Signal Process., pp. 179–204, 2018

    2018

  25. [34]

    Improving the readability of time-frequency and time-scale representations by the reassignment method,

    F. Auger and P. Flandrin, “Improving the readability of time-frequency and time-scale representations by the reassignment method,” IEEE Trans. on Signal Process. , vol. 43, no. 5, pp. 1068–1089, 1995

    1995

  26. [35]

    Time-frequency reassignment and synchrosqueezing: An overview,

    F. Auger, P. Flandrin, Y .-T. Lin, S. McLaughlin, S. Meignen, T. Oberlin, and H.-T. Wu, “Time-frequency reassignment and synchrosqueezing: An overview,” IEEE Signal Process. Magazine , vol. 30, no. 6, pp. 32–41, 2013

    2013

  27. [36]

    Synchrosqueezing-based recovery of in- stantaneous frequency from nonuniform samples,

    G. Thakur and H.-T. Wu, “Synchrosqueezing-based recovery of in- stantaneous frequency from nonuniform samples,” SIAM Journal on Mathematical Analysis, vol. 43, no. 5, pp. 2078–2095, 2011

    2011

  28. [37]

    The Fourier-based syn- chrosqueezing transform,

    T. Oberlin, S. Meignen, and V . Perrier, “The Fourier-based syn- chrosqueezing transform,” in 2014 IEEE int. conf. on acoustics, speech and Signal Process. (ICASSP) . IEEE, 2014, pp. 315–319

    2014

  29. [38]

    The synchrosqueez- ing algorithm for time-varying spectral analysis: Robustness properties and new paleoclimate applications,

    G. Thakur, E. Brevdo, N. S. Fu ˇckar, and H.-T. Wu, “The synchrosqueez- ing algorithm for time-varying spectral analysis: Robustness properties and new paleoclimate applications,” Signal Process., vol. 93, no. 5, pp. 1079–1094, 2013

    2013

  30. [39]

    Second-order synchrosqueezing transform or invertible reassignment? towards ideal time-frequency representations,

    T. Oberlin, S. Meignen, and V . Perrier, “Second-order synchrosqueezing transform or invertible reassignment? towards ideal time-frequency representations,” IEEE Trans. on Signal Process. , vol. 63, no. 5, pp. 1335–1344, 2015

    2015

  31. [40]

    Theoretical analysis of the second-order synchrosqueezing transform,

    R. Behera, S. Meignen, and T. Oberlin, “Theoretical analysis of the second-order synchrosqueezing transform,” Applied and Computational Harmonic Analysis, vol. 45, no. 2, pp. 379–404, 2018

    2018

  32. [41]

    Adaptive short-time Fourier transform and synchrosqueezing transform for non-stationary signal separation,

    L. Li, H. Cai, H. Han, Q. Jiang, and H. Ji, “Adaptive short-time Fourier transform and synchrosqueezing transform for non-stationary signal separation,” Signal Process., vol. 166, p. 107231, 2020

    2020

  33. [42]

    Recursive versions of the Levenberg-Marquardt reassigned spectrogram and of the syn- chrosqueezed STFT,

    D. Fourer, F. Auger, and P. Flandrin, “Recursive versions of the Levenberg-Marquardt reassigned spectrogram and of the syn- chrosqueezed STFT,” in 2016 IEEE Int. Conf. on Acoust., Speech and Signal Process. (ICASSP) . IEEE, 2016, pp. 4880–4884

    2016

  34. [43]

    Adaptive synchrosqueezing transform with a time-varying parameter for non-stationary signal separation,

    L. Li, H. Cai, and Q. Jiang, “Adaptive synchrosqueezing transform with a time-varying parameter for non-stationary signal separation,” Applied and Computational Harmonic Analysis , vol. 49, no. 3, pp. 1075–1106, 2020

    2020

  35. [44]

    Novel short-time fractional Fourier transform: Theory, implementation, and applications,

    J. Shi, J. Zheng, X. Liu, W. Xiang, and Q. Zhang, “Novel short-time fractional Fourier transform: Theory, implementation, and applications,” IEEE Trans. on Signal Process. , vol. 68, pp. 3280–3295, 2020

    2020

  36. [45]

    Synchrosqueezing-based short-time fractional Fourier transform,

    Z. Zhao and G. Li, “Synchrosqueezing-based short-time fractional Fourier transform,” IEEE Trans. on Signal Process. , vol. 71, pp. 279– 294, 2023

    2023

  37. [46]

    Backpropagation through time: what it does and how to do it,

    P. J. Werbos, “Backpropagation through time: what it does and how to do it,” Proceedings of the IEEE , vol. 78, no. 10, pp. 1550–1560, 1990

    1990

  38. [47]

    Measuring time-frequency information content using the R ´enyi entropies,

    R. G. Baraniuk, P. Flandrin, A. J. Janssen, and O. J. Michel, “Measuring time-frequency information content using the R ´enyi entropies,” IEEE Trans. on Information theory , vol. 47, no. 4, pp. 1391–1409, 2002

    2002

  39. [48]

    On the use of R ´enyi entropy for optimal window size computation in the short-time Fourier transform,

    S. Meignen, M. Colominas, and D. Pham, “On the use of R ´enyi entropy for optimal window size computation in the short-time Fourier transform,” in IEEE int. conf. on Acoustics, Speech and Signal Process. (ICASSP), 2020, pp. 5830–5834

    2020

  40. [49]

    Nonlocal operators with applications to image processing,

    G. Gilboa and S. Osher, “Nonlocal operators with applications to image processing,” Multiscale Modeling & Simulation, vol. 7, no. 3, pp. 1005– 1028, 2009

    2009

  41. [50]

    Speech emotion recognition from spectrograms with deep convolutional neural network,

    A. M. Badshah, J. Ahmad, N. Rahim, and S. W. Baik, “Speech emotion recognition from spectrograms with deep convolutional neural network,” in 2017 Int. Conf. Platform Technology and Service (PlatCon) . IEEE, 2017, pp. 1–5

    2017

Pith tools

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