Pith. sign in

REVIEW 4 major objections 6 minor 61 references

Quantized Approximate Signal Processing (QASP): Towards Homomorphic Encryption for audio

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Fully homomorphic encryption can compute standard audio time-frequency representations on encrypted raw audio.

desk verdict First FHE pipeline for raw-audio time-frequency features is real, but the abstract oversells the approximate STFT gains; the improvement tables are within statistical noise. read the letter →

arxiv 2505.10500 v1 pith:R7PJJTKA submitted 2025-05-15 eess.AS cs.CRcs.SD

classification eess.AScs.CRcs.SD
keywords fullyhomomorphicencryptionaudioprivacytime-frequencyrepresentationsSTFTquantizationMFCCgammatonefiltersapproximatesignalprocessing
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

This paper claims that standard audio front-ends—short-time Fourier transform, Mel filterbanks, Mel-frequency cepstral coefficients, and gammatone filters—can be computed entirely inside fully homomorphic encryption (FHE), directly on encrypted raw audio, without ever decrypting the signal. The authors cast these time-frequency representations as quantized neural-network layers with fixed weights, then compile them into FHE circuits under a 16-bit integer budget. They introduce several approximate STFT variants—dilation, frequency-dependent windows, poorman's projection, $\ell^1$ energy, and cropping—that reduce the bit cost of the computation and, for the best variants, lower the error of the encrypted outputs relative to conventional STFT. On VocalSet and OxVoc, the encrypted descriptors reproduce the statistical-test structure of clear computation with mean error rates between roughly 2% and 6%, and encrypted CNNs classify gender and vocal exercises with small accuracy drops. If correct, this means a server can run useful audio analytics on encrypted speech and return results to a client without ever seeing the plaintext audio.

What carries the argument

The load-bearing object is the quantized approximate STFT, treated as a fixed-weight convolutional layer whose input, kernel weights, outputs, and intermediate activations each have their own bit width, constrained so that no intermediate exceeds 16 bits. The argument runs through the bit-growth formula for a dot product, which shows that sparsifying the kernel (dilation, cropping), narrowing windows (frequency-dependent windows), or replacing complex exponentials by projections onto the $L$-th roots of unity (poorman's transform) reduces the bits needed for accumulation; the freed budget is then spent on higher input/output precision, which matters more for downstream statistical tests and classifiers. The poorman's transform comes with a provable error bound $\|X-X_L\|\le 2|\sin(\pi/(2L))|\sqrt{\sum_i (x(i)w(i-mh))^2}$, and the dilation error is expressed as leakage from other frequency bins. These approximations are not free: they add their own approximation error, and the paper's heuristic is that the quantization-error reduction outweighs it.

What would settle it

Take the exact grid-selected quantization parameters and bit widths reported for the best conventional and approximate models on VocalSet, and evaluate them on a held-out set of the same classes recorded in a different room, microphone, or signal-to-noise ratio; if the statistical-test error rate jumps well above the reported 1.9-6.5% range, or encrypted classification accuracy drops far below the clear baseline, then the calibration-to-evaluation generalization on which the results depend has failed. A sharper test of the approximation claim: measure the FHE-vs-clear distance for a signal whose energy lies mostly above 4 kHz, where the paper states quantization zeroes the spectrum; a large distance there is expected, and only matters if downstream tasks rely on that band.

Watch

Extended reading notes

Core claim

The central discovery is that the computational bottleneck of FHE audio processing is quantization error, not cryptographic noise, and that this error can be steered by choosing where bits are spent. By writing the STFT as a convolution, all four representations become quantized neural-network operations, and the accumulation width grows as $\lceil \log_2(L(2^N-1)(2^M-1))\rceil$ for a length-$L$ dot product with $N$- and $M$-bit operands. The paper shows that inserting zeros into the STFT kernel (dilation, cropping), adapting windows to frequency, or projecting complex exponentials onto a finite set of roots of unity (poorman's transform, $L=4$) lowers the bit demand and lets more bits go to inputs and activations, yielding encrypted outputs that are closer to the clear computation than the conventional STFT is in FHE. Empirically, the best approximations cut mean statistical-test error rates from 5.9% to 4.7% on VocalSet and from 6.5% to 1.9% on OxVoc, and encrypted CNNs classify gender near clear accuracy while vocal-exercise classification stays within a few points of clear. The paper also acknowledges a cost: quantization zeroes much of the energy above 4 kHz in the spectrograms.

Load-bearing premise

The pipeline rests on the assumption that quantization parameters (bit widths, and min/max calibration ranges) chosen on a stratified 10% calibration set remain good for the other 90% of the data; if the calibration audio is not representative of the evaluation distribution, the reported FHE-vs-clear errors are optimistic.

Editorial extensions

If this is right

  • A server can compute STFT, Mel filterbanks, MFCCs, and gammatone features on encrypted raw audio and return them, or their statistics, to the client, enabling private speech analytics without exposing the signal.
  • Approximation is not merely a lossy shortcut: in FHE, approximate STFT variants such as dilation and poorman can be more faithful to the clear computation than the conventional STFT, because they reduce quantization error.
  • End-to-end encrypted classification from raw audio is feasible for at least simple CNNs: gender classification stays within about 0.05 of clear accuracy, and 4-class vocal exercise classification remains well above chance.
  • A practical heuristic for FHE parameter selection follows from the bit-budget analysis: spend more bits on activations and outputs, fewer on convolution weights, which the post-hoc bit distributions support.

Reading between the lines

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

  • The same descriptor-based pipeline could be pointed at clinical or paralinguistic biomarkers (for example, vocal jitter or shimmer-like statistics) without architectural change, since the paper already computes mean and standard-deviation statistics over Mel and gammatone energies; a direct test is whether the low error rates survive on disordered speech.
  • Because the poorman and dilation error bounds are explicit functions of $L$ and $d$, one could replace the grid search for bit widths with a principled, bound-driven parameter selection, treating the allowed approximation error as a budget to be split between quantization and approximation.
  • The reported runtime of about 12,970 seconds for 64 ms of audio suggests that the practical bottleneck has shifted from accuracy to speed; the approximate kernels' sparsity may cut the number of encrypted multiplications and thus latency, an effect the paper does not quantify.
  • The large intrinsic distance for MFCCs hints that cepstral features need their own approximation strategy, for example handling the log and discrete cosine transform stages separately, before FHE-MFCCs are used in applications that depend on fine cepstral structure.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper presents Quantized Approximate Signal Processing (QASP), a framework for computing standard time-frequency representations (STFT, Mel filterbanks, MFCCs, gammatone filters) on encrypted raw audio under the TFHE scheme, using quantized neural network layers and several approximate STFT formulations (dilation, frequency-dependent windows, poorman's projection, l1 energy, cropping). The authors derive error bounds for the approximations and evaluate the pipeline on VocalSet and OxVoc through intrinsic spectrogram distances, statistical tests over audio descriptors, and CNN gender/vocal-exercise classification, comparing conventional STFT with the approximations. The paper claims that the approximations significantly improve error rates and classification performance relative to conventional FHE STFT.

Significance. If the empirical claims were fully supported, this would be a valuable proof-of-concept: it would show that a server can compute standard audio features and simple classifiers over encrypted raw audio without seeing plaintext, with error bounds guiding parameter choice. The derivation of the poorman's transform error bound (Appendix A.2) and the decomposition of quantization bit widths (Section 3.3) are clear strengths, as are the use of public datasets and the explicit statement of limitations (lost harmonics above 4 kHz, very high FHE latency). However, the headline claims about approximation gains are not established by the reported numbers, and several statements go beyond what the tables show.

major comments (4)
  1. [Section 5.2, Tables 2 and 3; Abstract] The claim that the approximations 'substantially reduce error rates' and yield 'significant performance improvements' is not supported by the reported error counts. The largest improvement over the conventional STFT in any cell is 3 errors (e.g., 15 vs 12 on VocalSet Mean Gammatone; 3 vs 1 on OxVoc Std RMS). For paired binary outcomes, a McNemar exact test with a total-error difference of 3 has a minimum two-sided p-value of 0.25, and differences of 1 or 2 have p-values of at least 0.5; therefore no cell in these tables can reach statistical significance no matter how the discordant pairs are arranged. The paper reports no significance test, no confidence intervals, and no raw confusion matrices. Moreover, the best variant is selected post hoc among 12 approximations, and the winner changes across markers and datasets (e.g., OxVoc Mean Mel has 0.0 errors for Poorman L=6, Poorman L=8, and Dilation d=2, while Dilation d=4 gives 3.7), so the inference that 'dilation factor 4 ... is the best approach across both datasets' is not established.
  2. [Sections 4.1.2 and 4.2] The quantization bit widths Bi, Bo, Bw, and Bm are selected by grid search on a 10% calibration set and then used to evaluate on the remaining 90%, but the manuscript provides no evidence that the selected widths generalize to the evaluation distribution. Because the calibration criterion is Pearson correlation with clear descriptors, the comparison of approximation methods on the evaluation set can be biased toward methods that happen to overfit the calibration split. Please report repeated calibration/evaluation splits (or a bootstrap) and the stability of the selected bit widths, or otherwise demonstrate that the ranking of approximations is robust to the calibration set choice.
  3. [Section 5.3, Tables 4 and 5] The statement that the approximations 'helped to get closer to the clear computation' for CNN classification is inconsistent with the tables. For binary gender classification, Poorman is worse than the conventional FHE pipeline for every representation (e.g., 0.82 vs 0.89 for STFT) and Dilation is worse for Mel (0.82 vs 0.86) and Gammatone (0.82 vs 0.89). For 4-class vocal exercises, Dilation improves only STFT and Mel, and Poorman degrades STFT (0.50 vs 0.54) and MFCC (0.50 vs 0.53) relative to the conventional STFT. No significance test is reported across the 5 runs, and the standard deviations (0.01–0.04) are large relative to the differences, so the claimed classification improvements are not established.
  4. [Section 1 and Highlights] The 'first' claims (first system to perform FHE on raw audio; 'First complete end-to-end demonstration of FHE applied directly to raw audio signals') are not reconciled with the authors' own description of Zhang et al. [59] as avoiding filterbanks and 'using only smaller convolutions on audio'. If prior work already applied homomorphic computations directly to raw audio, the novelty should be stated more narrowly (e.g., first to compute these four time-frequency representations under TFHE, or first with this particular CNN pipeline). Please cite and compare explicitly so that the contribution is accurately scoped.
minor comments (6)
  1. [Equation (7) and Appendix A.1] The dilation indicator function notation is ambiguous: '1_{n≡0[min(d,dk)]}' should be defined as n ≡ 0 mod min(d, d_k), and Appendix A.1 uses a different form '1_{n≡0[d]}'; please align the notation and define all symbols.
  2. [Figure 3] The caption says the figure shows the FHE-friendly transformation 'corresponding to lowest Euclidean distance' but does not specify which approximation is used; please state the configuration explicitly.
  3. [Section 5.1 and Section 6] The conclusion states that all four time-frequency representations were 'successfully demonstrated', but Section 5.1 reports MFCC mean normalized 2D Euclidean distances of 0.78–0.85 with maxima above 1.5, far above the other features (0.13–0.18); please qualify the MFCC claim and discuss the practical utility of the encrypted MFCCs.
  4. [Section 5.5] The limitation section reports an FHE latency of 12,970 seconds for a 64 ms audio clip; this practical constraint should be reflected in the abstract or conclusion to avoid a misleading impression about deployability.
  5. [General] The paper does not include a data/code availability statement; given the many pipeline parameters, releasing code or a detailed configuration file would improve reproducibility.
  6. [General] There are minor typographical inconsistencies (e.g., 'Vocalset' vs 'VocalSet' in Figure 3, a stray period in the Table 3 caption); please proofread carefully.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the FHE pipeline, quantization framework, and approximation error bounds are derived from external tools and stated mathematical assumptions, with self-citations only in contextual passages.

full rationale

The paper's derivation chain is self-contained. The secure audio pipeline is built on external primitives (TFHE/Concrete, Brevitas) and standard signal-processing identities; the STFT is expressed as a fixed-weight convolution, and the Mel/MFCC/gammatone transforms are likewise fixed-weight layers. Each approximate STFT variant is introduced by an explicit defining equation (e.g., Eq. 7 for dilation, Eq. 9 for poorman's transform) and accompanied by error bounds derived in the text, such as the Cauchy-Schwarz bound in Appendix A.2. No fitted parameter is relabeled as a prediction: bit-width parameters are selected by grid search on a 10% calibration split and evaluated on the held-out 90%, and the claim that dilation d=4 performs best is an empirical comparison across variants, not a quantity forced by construction. The two author self-citations, Riad et al. 2020 and 2022, appear only as background for clinical speech markers and are not used to justify any methodological step. The absence of significance testing and the post-hoc choice of the winning approximation are statistical-validity concerns rather than circularity, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or cryptographic entities. The free parameters are mostly hyperparameters of the approximations and quantization scheme, which are fitted to data or chosen by hand. The main assumptions are the honest-client model, the 16-bit Concrete constraint, and the dominance of quantization error over compilation error.

free parameters (6)
  • Quantization bit widths (Bi, Bo, Bw, Bm) = grid-searched per experiment
    Four bit-width parameters control input, output, weight, and intermediate quantization; selected to maximize Pearson correlation with clear descriptors or classification accuracy on calibration data.
  • Dilation factor d = 2, 3, 4, 5, or d_k, chosen per experiment
    The dilation approximation skips samples in the STFT kernel; the best d was selected post hoc after evaluating multiple values.
  • Poorman's L = 4, 6, or 8
    Number of quantized angles for the poorman's transform; larger L gives finer frequency quantization but uses more bits.
  • N_min for frequency-adapted windows = 80
    Minimum window width at the highest frequency bin; chosen by hand for the frequency-dependent window approximation.
  • Cropping frequency band = f_min=0 Hz, f_max=1000 Hz
    STFT coefficients outside this range are set to zero; the heuristic assumes voice information concentrates below 1 kHz.
  • STFT window length N and hop length h = not stated explicitly
    Standard STFT parameters required for the computations; their values are not reported in the text, affecting reproducibility.
assumptions (5)
  • ad hoc to paper The FHE compilation error is negligible compared to the quantization error.
    Section 3, first paragraph: 'The compilation error can be neglected in comparison to the quantization error.' This premise justifies focusing the analysis on quantization, but the paper provides no direct measurement of compilation noise.
  • domain assumption The client is honest.
    Section 2.2: 'We assume in this paper that the client is honest.' This is a standard cryptographic assumption but limits the security claim to honest-client settings.
  • domain assumption Concrete enforces a 16-bit limit over all intermediate values.
    Section 2.2: 'Concrete also enforces a 16-bit limit over all intermediate values of the circuit in order to keep the FHE duration low.' The entire bit-optimization strategy depends on this tooling constraint.
  • standard math Shannon-Nyquist sampling theorem can determine the maximum safe dilation rate per frequency bin.
    Section 3.4 uses the theorem to derive d_k = floor(N/(2(k+1))). This assumes each frequency bin is approximately bandlimited to its upper edge, which is only a heuristic for short-window audio.
  • standard math Cauchy-Schwarz inequality for the poorman's transform error bound.
    Appendix A.2 uses Cauchy-Schwarz to bound the difference between the true and poorman's DFT; this is a valid mathematical step.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantized Approximate Signal Processing (QASP): Towards Homomorphic Encryption for audio." pith.science (2026). https://pith.science/paper/R7PJJTKA

@misc{pith2026250510500,
  author       = {Pith},
  title        = {Pith review of: Quantized Approximate Signal Processing (QASP): Towards Homomorphic Encryption for audio},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7PJJTKA}},
  note         = {Machine review of arXiv:2505.10500}
}
read the original abstract

Audio and speech data are increasingly used in machine learning applications such as speech recognition, speaker identification, and mental health monitoring. However, the passive collection of this data by audio listening devices raises significant privacy concerns. Fully homomorphic encryption (FHE) offers a promising solution by enabling computations on encrypted data and preserving user privacy. Despite its potential, prior attempts to apply FHE to audio processing have faced challenges, particularly in securely computing time frequency representations, a critical step in many audio tasks. Here, we addressed this gap by introducing a fully secure pipeline that computes, with FHE and quantized neural network operations, four fundamental time-frequency representations: Short-Time Fourier Transform (STFT), Mel filterbanks, Mel-frequency cepstral coefficients (MFCCs), and gammatone filters. Our methods also support the private computation of audio descriptors and convolutional neural network (CNN) classifiers. Besides, we proposed approximate STFT algorithms that lighten computation and bit use for statistical and machine learning analyses. We ran experiments on the VocalSet and OxVoc datasets demonstrating the fully private computation of our approach. We showed significant performance improvements with STFT approximation in private statistical analysis of audio markers, and for vocal exercise classification with CNNs. Our results reveal that our approximations substantially reduce error rates compared to conventional STFT implementations in FHE. We also demonstrated a fully private classification based on the raw audio for gender and vocal exercise classification. Finally, we provided a practical heuristic for parameter selection, making quantized approximate signal processing accessible to researchers and practitioners aiming to protect sensitive audio data.

Figures

Figures reproduced from arXiv: 2505.10500 by the authors.

Figure 1
Figure 1. We assume in this paper that the client is honest Bäckström [4]. [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 1
Figure 1. Schematic representation of secure audio processing with fully homomorphic [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. 9 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figures from the paper (4 more)
Figure 2
Figure 2. Figure 2: Top: conventional computation of the spectrogram, or squared magnitude of [PITH_FULL_IMAGE:figures/full_fig_p013_2.png]
Figure 3
Figure 3. Figure 3: Comparison of log-scale spectrograms with FHE-friendly transformation corre [PITH_FULL_IMAGE:figures/full_fig_p014_3.png]
Figure 4
Figure 4. Figure 4: Example of a scatter plot of p-values computed in clear or in FHE with the [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Distributions of bit width controlling the input quantization ( [PITH_FULL_IMAGE:figures/full_fig_p026_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [59]

    Encrypted speech recog- nition using deep polynomial networks

    Shi-Xiong Zhang, Yifan Gong, and Dong Yu. Encrypted speech recog- nition using deep polynomial networks. InICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5691–5695. IEEE, 2019. 32 Appendix A. STFT approximations Appendix A.1. Calculations for dilation X (d)(m,k) = X l,ld<N x[ld]w[ld−mh]e −2iπkld N...

  2. [1]

    Heir: Homomorphic encryption intermediate representation.https: //heir.dev/

  3. [2]

    Abdul and Abdulbasit K

    Zrar Kh. Abdul and Abdulbasit K. Al-Talabani. Mel frequency cepstral coefficient and its applications: A review.IEEE Access, 10:122136– 122158, 2022. doi: 10.1109/ACCESS.2022.3223444

  4. [3]

    Speech emotion recognition: Emotional models, databases, features, preprocessing methods, support- ing modalities, and classifiers.Speech Communication, 116:56–76, 2020

    Mehmet Berkehan Akçay and Kaya Oğuz. Speech emotion recognition: Emotional models, databases, features, preprocessing methods, support- ing modalities, and classifiers.Speech Communication, 116:56–76, 2020

  5. [4]

    Privacy in speech technology.arXiv preprint arXiv:2305.05227, 2023

    Tom Bäckström. Privacy in speech technology.arXiv preprint arXiv:2305.05227, 2023

  6. [5]

    an efficient algorithm for the cal- culation of a constant q transform

    Judith Brown and Miller Puckette. "an efficient algorithm for the cal- culation of a constant q transform".Journal of the Acoustical Society of America, 92:2698, 11 1992. doi: 10.1121/1.404385

  7. [6]

    Homomor- phic encryption for arithmetic of approximate numbers

    Jung Cheon, Andrey Kim, Miran Kim, and Yongsoo Song. Homomor- phic encryption for arithmetic of approximate numbers. pages 409–437, 11 2017. ISBN 978-3-319-70693-1. doi: 10.1007/978-3-319-70694-8_15

  8. [7]

    nnaudio: An on-the-fly gpu audio to spectrogram conversion toolbox using 1d convolutional neural networks.IEEE Access, 8:161981–162003,

    Kin Wai Cheuk, Hans Anderson, Kat Agres, and Dorien Herremans. nnaudio: An on-the-fly gpu audio to spectrogram conversion toolbox using 1d convolutional neural networks.IEEE Access, 8:161981–162003,

Show all 61 references
  1. [8]

    Tfhe: Fast fully homomorphic encryption over the torus

    Ilaria Chillotti, Nicolas Gama, Mariya Georgieva, and Malika Iz- abachène. Tfhe: Fast fully homomorphic encryption over the torus. Cryptology ePrint Archive, Paper 2018/421, 2018. URLhttps:// eprint.iacr.org/2018/421.https://eprint.iacr.org/2018/421

  2. [9]

    Voxceleb2: Deep speaker recognition

    Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. InInterspeech, 2018. URLhttps://api. semanticscholar.org/CorpusID:49211906. 26

  3. [10]

    A review of depression and suicide risk assessment using speech analysis.Speech communica- tion, 71:10–49, 2015

    Nicholas Cummins, Stefan Scherer, Jarek Krajewski, Sebastian Schnieder, Julien Epps, and Thomas F Quatieri. A review of depression and suicide risk assessment using speech analysis.Speech communica- tion, 71:10–49, 2015

  4. [11]

    Sine-wave speech and privacy-preserving depression detec- tion

    Sri Harsha Dumpala, Rudolf Uher, Stan Matwin, Michael Kiefte, and Sageev Oore. Sine-wave speech and privacy-preserving depression detec- tion. InProc. SMM21, Workshop on Speech, Music and Mind, volume 2021, pages 11–15, 2021

  5. [12]

    Opensmile: the munich versatile and fast open-source audio feature extractor

    Florian Eyben, Martin Wöllmer, and Björn Schuller. Opensmile: the munich versatile and fast open-source audio feature extractor. InPro- ceedings of the 18th ACM international conference on Multimedia, pages 1459–1462, 2010

  6. [13]

    Linguistic features identify alzheimer’s disease in narrative speech.Journal of Alzheimer’s Disease, 49(2):407–422, 2016

    Kathleen C Fraser, Jed A Meltzer, and Frank Rudzicz. Linguistic features identify alzheimer’s disease in narrative speech.Journal of Alzheimer’s Disease, 49(2):407–422, 2016

  7. [14]

    Privacy- preservingtree-basedinferencewithtfhe

    Jordan Frery, Andrei Stoian, Roman Bredehoft, Luis Montero, Celia Kherfallah, Benoit Chevallier-Mames, and Arthur Meyre. Privacy- preservingtree-basedinferencewithtfhe. InInternational Conference on Mobile, Secure, and Programmable Networking, pages139–156.Springer, 2023

  8. [15]

    Optimization of rnn-based speech activitydetection.IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, PP:1–1, 11 2017

    Gregory Gelly and Jean-Luc Gauvain. Optimization of rnn-based speech activitydetection.IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, PP:1–1, 11 2017. doi: 10.1109/TASLP.2017.2769220

  9. [16]

    Fully homomorphic encryption using ideal lattices

    Craig Gentry. Fully homomorphic encryption using ideal lattices. In Proceedings of the forty-first annual ACM symposium on Theory of com- puting, pages 169–178, 2009

  10. [17]

    Mood state prediction from speech of varying acoustic quality for individuals with bipolar disorder

    John Gideon, Emily Mower Provost, and Melvin McInnis. Mood state prediction from speech of varying acoustic quality for individuals with bipolar disorder. In2016 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 2359–2363. IEEE, 2016

  11. [18]

    Privacy preserving encrypted phonetic search of speech data

    Cornelius Glackin, Gerard Chollet, Nazim Dugan, Nigel Cannings, Julie Wall, Shahzaib Tahir, Indranil Ghosh Ray, and Muttukrishnan Rajara- jan. Privacy preserving encrypted phonetic search of speech data. In 27 2017 IEEE International Conference on Acoustics, Speech and Signal ...

  12. [19]

    AST: Audio Spectrogram Transformer

    Yuan Gong, Yu-An Chung, and James Glass. AST: Audio Spectrogram Transformer. InProc. Interspeech 2021, pages 571–575, 2021. doi: 10. 21437/Interspeech.2021-698

  13. [20]

    Birkhäuser Boston, 2001

    Karlheinz Gröchenig.The Short-Time Fourier Transform, pages 37–58. Birkhäuser Boston, 2001. doi: 10.1007/978-1-4612-0003-1_4

  14. [21]

    Training speech recognition models with federated learning: A quality/cost framework

    Dhruv Guliani, Françoise Beaufays, and Giovanni Motta. Training speech recognition models with federated learning: A quality/cost framework. InICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3080–3084. IEEE, 2021

  15. [22]

    Voxceleb enrichment for age and gender recognition

    Khaled Hechmi, Trung Ngo Trong, Ville Hautamäki, and Tomi Kin- nunen. Voxceleb enrichment for age and gender recognition. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 687–693. IEEE, 2021

  16. [23]

    Measuring relative attack surfaces.Computer Security in the 21st Cen- tury, page 109, 2005

    Michael Howard, WA Redmond, Jon Pincus, and Jeannette M Wing. Measuring relative attack surfaces.Computer Security in the 21st Cen- tury, page 109, 2005

  17. [24]

    Advances and open prob- lems in federated learning.Foundations and trends®in machine learn- ing, 14(1–2):1–210, 2021

    Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, MehdiBennis, ArjunNitinBhagoji, KallistaBonawitz, ZacharyCharles, Graham Cormode, Rachel Cummings, et al. Advances and open prob- lems in federated learning.Foundations and trends®in machine learn- ing, 14(1–2):...

  18. [25]

    Quantizing deep convolutional networks for efficient inference: A whitepaper.ArXiv, abs/1806.08342:4–5, 2018

    Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper.ArXiv, abs/1806.08342:4–5, 2018. URLhttps://api.semanticscholar.org/CorpusID:49356451

  19. [26]

    Lamoureux

    M.P. Lamoureux. The poorman’s transform: approximating the fourier transform without multiplication.IEEE Transactions on Signal Pro- cessing, 41(3):1413–1415, 1993. doi: 10.1109/78.205744

  20. [27]

    Exploring the granularity of sparsity in convolutional 28 neural networks

    Huizi Mao, Song Han, Jeff Pool, Wenshuo Li, Xingyu Liu, Yu Wang, and William J Dally. Exploring the granularity of sparsity in convolutional 28 neural networks. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 13–20, 2017

  21. [28]

    Convolution-friendly image compression in FHE

    Axel Mertens, Georgio Nicolas, and Sergi Rovira. Convolution-friendly image compression in FHE. Cryptology ePrint Archive, Paper 2024/559,

  22. [29]

    Estimation of unknown speaker’s height from speech.International Journal of Speech Technology, 12: 149–160, 2009

    Iosif Mporas and Todor Ganchev. Estimation of unknown speaker’s height from speech.International Journal of Speech Technology, 12: 149–160, 2009

  23. [30]

    Behavioral signal processing: Deriving human behavioral informatics from speech and lan- guage.Proceedings of the IEEE, 101(5):1203–1233, 2013

    Shrikanth Narayanan and Panayiotis G Georgiou. Behavioral signal processing: Deriving human behavioral informatics from speech and lan- guage.Proceedings of the IEEE, 101(5):1203–1233, 2013

  24. [31]

    Homomorphic encryption for speaker recognition: Protection of biometric templates and vendor model pa- rameters.arXiv preprint arXiv:1803.03559, 2018

    Andreas Nautsch, Sergey Isadskiy, Jascha Kolberg, Marta Gomez- Barrero, and Christoph Busch. Homomorphic encryption for speaker recognition: Protection of biometric templates and vendor model pa- rameters.arXiv preprint arXiv:1803.03559, 2018

  25. [32]

    Preserving privacy in speaker and speech characterisation.Computer Speech & Language, 58: 441–480, 2019

    Andreas Nautsch, Abelino Jiménez, Amos Treiber, Jascha Kolberg, Catherine Jasserand, Els Kindt, Héctor Delgado, Massimiliano Todisco, Mohamed Amine Hmani, Aymen Mtibaa, et al. Preserving privacy in speaker and speech characterisation.Computer Speech & Language, 58: 441–480, 2019

  26. [33]

    Privacy-preserving speaker recognition with cohort score normalisation

    Andreas Nautsch, Jose Patino, Amos Treiber, Themos Stafylakis, Petr Mizera, Massimiliano Todisco, Thomas Schneider, and Nicholas Evans. Privacy-preserving speaker recognition with cohort score normalisation. InInterspeech 2019, pages 2868–2872. ISCA, 2019

  27. [34]

    Xilinx/brevitas, 2023

    Alessandro Pappalardo. Xilinx/brevitas, 2023. URLhttps://doi.org/ 10.5281/zenodo.3333552

  28. [35]

    Parsons, Katherine S

    Christine E. Parsons, Katherine S. Young, Michelle G. Craske, Alan L. Stein, and Morten L. Kringelbach. Introducing the oxford vocal (oxvoc) sounds database: a validated set of non-acted affective sounds from human infants, adults, and domestic animals.Frontiers in Psychology,...

  29. [36]

    Privacy preserving probabilistic inference with hidden markov models

    Manas Pathak, Shantanu Rane, Wei Sun, and Bhiksha Raj. Privacy preserving probabilistic inference with hidden markov models. In2011 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 5868–5871. IEEE, 2011

  30. [37]

    Privacy-preserving speaker verifica- tion and identification using gaussian mixture models.IEEE Transac- tions on Audio, Speech, and Language Processing, 21(2):397–406, 2012

    Manas A Pathak and Bhiksha Raj. Privacy-preserving speaker verifica- tion and identification using gaussian mixture models.IEEE Transac- tions on Audio, Speech, and Language Processing, 21(2):397–406, 2012

  31. [38]

    Anefficient auditory filterbank based on the gammatone function

    R.Patterson, IanNimmo-Smith, J.Holdsworth, andP.Rice. Anefficient auditory filterbank based on the gammatone function. 01 1988

  32. [39]

    Federated learning with differential privacy for end-to-end speech recognition.arXiv preprint arXiv:2310.00098, 2023

    Martin Pelikan, Sheikh Shams Azam, Vitaly Feldman, Jan Silovsky, Kunal Talwar, Tatiana Likhomanenko, et al. Federated learning with differential privacy for end-to-end speech recognition.arXiv preprint arXiv:2310.00098, 2023

  33. [40]

    Vocal-source biomarkers for depression: A link to psychomotor activity

    Thomas F Quatieri and Nicolas Malyska. Vocal-source biomarkers for depression: A link to psychomotor activity. InInterspeech, volume 2, pages 1059–1062, 2012

  34. [41]

    Robust speech recognition via large- scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large- scale weak supervision. InInternational conference on machine learning, pages 28492–28518. PMLR, 2023

  35. [42]

    Evaluation of the pain level from speech: Introducing a novel pain database and benchmarks

    Zhao Ren, Nicholas Cummins, Jing Han, Sebastian Schnieder, Jarek Krajewski, and Bjorn Schuller. Evaluation of the pain level from speech: Introducing a novel pain database and benchmarks. InSpeech Commu- nication; 13th ITG-Symposium, pages 1–5. VDE, 2018

  36. [43]

    Vocal markers from sustained phonation in huntington’s disease

    Rachid Riad, Hadrien Titeux, Laurie Lemoine, Justine Montillot, Jen- nifer Hamet Bagnou, Xuan Nga Cao, Emmanuel Dupoux, and Anne- Catherine Bachoud-Lévi. Vocal markers from sustained phonation in huntington’s disease. InINTERSPEECH 2020-Annual Conference of the International S...

  37. [44]

    Predicting clinical scores in huntington’s disease: a lightweight speech test.Journal of neurology, 269(9):5008–5021, 2022

    Rachid Riad, Marine Lunven, Hadrien Titeux, Xuan-Nga Cao, Jennifer Hamet Bagnou, Laurie Lemoine, Justine Montillot, Agnes Sliwinski, Katia Youssov, Laurent Cleret de Langavant, et al. Predicting clinical scores in huntington’s disease: a lightweight speech test.Journal of neur...

  38. [45]

    A survey of privacy attacks in machine learning.ACM Computing Surveys, 56(4):1–34, 2023

    Maria Rigaki and Sebastian Garcia. A survey of privacy attacks in machine learning.ACM Computing Surveys, 56(4):1–34, 2023

  39. [46]

    Speech emotion recognition: Two decades in a nut- shell, benchmarks, and ongoing trends.Communications of the ACM, 61(5):90–99, 2018

    Björn W Schuller. Speech emotion recognition: Two decades in a nut- shell, benchmarks, and ongoing trends.Communications of the ACM, 61(5):90–99, 2018

  40. [47]

    Differentially private speaker anonymiza- tion.Proceedings on Privacy Enhancing Technologies, 1:98–114, 2023

    Ali Shahin Shamsabadi, Brij Mohan Lal Srivastava, Aurélien Bellet, Nathalie Vauquier, Emmanuel Vincent, Mohamed Maouche, Marc Tom- masi, and Nicolas Papernot. Differentially private speaker anonymiza- tion.Proceedings on Privacy Enhancing Technologies, 1:98–114, 2023

  41. [48]

    X-vectors: Robust dnn embeddings for speaker recognition.2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5329–5333, 2018

    David Snyder, Daniel Garcia-Romero, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur. X-vectors: Robust dnn embeddings for speaker recognition.2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5329–5333, 2018. URLhttps: //api.semant...

  42. [49]

    Privacy and utility of x-vector based speaker anonymization.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:2383–2395, 2022

    Brij Mohan Lal Srivastava, Mohamed Maouche, Md Sahidullah, Em- manuel Vincent, Aurélien Bellet, Marc Tommasi, Natalia Tomashenko, Xin Wang, and Junichi Yamagishi. Privacy and utility of x-vector based speaker anonymization.IEEE/ACM Transactions on Audio, Speech, and Language P...

  43. [50]

    Deep neural networks for encrypted inference with tfhe

    Andrei Stoian, Jordan Frery, Roman Bredehoft, Luis Montero, Celia Kherfallah, and Benoit Chevallier-Mames. Deep neural networks for encrypted inference with tfhe. InInternational Symposium on Cyber Security, Cryptology, and Machine Learning, pages 493–500. Springer, 2023

  44. [51]

    Deep neural networks for en- crypted inference with tfhe

    Andrei Stoian, Jordan Frery, Roman Bredehoft, Luis Montero, Celia Kherfallah, and Benoit Chevallier-Mames. Deep neural networks for en- crypted inference with tfhe. In Shlomi Dolev, Ehud Gudes, and Pas- cal Paillier, editors,Cyber Security, Cryptology, and Machine Learning, pa...

  45. [52]

    Privacy-preserving automatic speaker diarization

    Francisco Teixeira, Alberto Abad, Bhiksha Raj, and Isabel Trancoso. Privacy-preserving automatic speaker diarization. InICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

  46. [53]

    Improving membership inference in asr model auditing with perturbed loss features.arXiv preprint arXiv:2405.01207, 2024

    Francisco Teixeira, Karla Pizzi, Raphael Olivier, Alberto Abad, Bhiksha Raj, and Isabel Trancoso. Improving membership inference in asr model auditing with perturbed loss features.arXiv preprint arXiv:2405.01207, 2024

  47. [54]

    Extracting mel-frequency and bark- frequency cepstral coefficients from encrypted signals

    Patricia Thaine and Gerald Penn. Extracting mel-frequency and bark- frequency cepstral coefficients from encrypted signals. InInterspeech, pages 3715–3719, 2019

  48. [55]

    Privacy-preserving plda speaker verification us- ing outsourced secure computation.Speech Communication, 114:60–71, 2019

    Amos Treiber, Andreas Nautsch, Jascha Kolberg, Thomas Schneider, and Christoph Busch. Privacy-preserving plda speaker verification us- ing outsourced secure computation.Speech Communication, 114:60–71, 2019

  49. [56]

    Vocalset: A singing voice dataset

    Julia Wilkins, Prem Seetharaman, Alison Wahl, and Bryan Pardo. Vocalset: A singing voice dataset. InInternational Society for Mu- sic Information Retrieval Conference, 2018. URLhttps://api. semanticscholar.org/CorpusID:53875542

  50. [57]

    Multi-scale context aggregation by di- lated convolutions, 2016

    Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by di- lated convolutions, 2016. URLhttps://arxiv.org/abs/1511.07122

  51. [58]

    Concrete: TFHE Compiler that converts python programs into FHE equivalent, 2022.https://github.com/zama-ai/concrete

    Zama. Concrete: TFHE Compiler that converts python programs into FHE equivalent, 2022.https://github.com/zama-ai/concrete

  52. [2020]

    doi: 10.1109/ACCESS.2020.3019084

  53. [2024]

    iacr.org/2024/559

    URLhttps://eprint.iacr.org/2024/559.https://eprint. iacr.org/2024/559

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.