Pith. sign in

REVIEW 3 major objections 6 minor 80 references

Gender bias in audio deepfake detectors is set by who you train on, and no threshold fix can close the core error gap.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 14:47 UTC pith:5G6US73P

load-bearing objection Controlled composition experiment that cleanly shows bias direction tracks the minority gender and that Oracle calibration cannot move the EER gap; speaker-overlap is a real but non-load-bearing limit. the 3 major comments →

arxiv 2607.09891 v1 pith:5G6US73P submitted 2026-07-10 cs.SD cs.AI

What You Train Is What You Get: Gender Bias, Training Composition, and Post-Hoc Mitigation in Audio Deepfake Detection

Aishwarya R. Fursule , Vamshi Nallaguntla , Shruti Kshirsagar , Anderson R. Avila This is my paper
classification cs.SD cs.AI
keywords audio deepfake detectiongender biastraining data compositionthreshold calibrationASVspoof5WavLMEqual Error Rate gapgroup fairness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that gender bias in audio deepfake detection is largely written by training composition: the gender that is scarce in training is the one that fails more at test time, across nearly all attacks. Using controlled custom splits of ASVspoof5, the authors train hundreds of attack-specific ResNet18 models under nine gender mixes, from female-only to male-only, with both handcrafted LogSpectrogram features and pretrained WavLM-Base+ embeddings. Composition predicts bias direction; WavLM gaps are several times larger than LogSpectrogram gaps under the same data; balanced training nearly erases LogSpectrogram disparity but leaves WavLM largely biased. All six post-hoc threshold calibrations, including an Oracle that peeks at test labels, leave the Equal Error Rate gap fixed at 1.317 percentage points. The practical claim is that fairness has to be built into training data and representation choice, because recalibration can only move the operating point, not reshape the score distributions that produce the gap.

Core claim

Training gender composition strongly predicts which gender is disadvantaged at test time, and the Equal Error Rate gap between genders is invariant to post-hoc threshold calibration—including Oracle calibration with full evaluation labels—remaining fixed at 1.317 pp across nine compositions. Feature type modulates magnitude: WavLM-Base+ produces gaps 3.0–4.3× larger than LogSpectrogram on identical training sets, and balanced data nearly eliminates LogSpectrogram bias while leaving WavLM bias largely intact.

What carries the argument

Nine controlled gender-composition training sets (female-only through male-only, plus partial mixes) on custom per-attack ASVspoof5 splits, paired with six threshold strategies including Oracle, used to separate composition effects from post-hoc operating-point choice and to show that EER gap tracks score-distribution disparity rather than threshold placement.

Load-bearing premise

The custom 70/10/20 splits are treated as cleanly isolating gender-composition effects even though they are not guaranteed to keep speakers out of both train and test, so some of the measured bias may reflect speaker reuse rather than composition alone.

What would settle it

Retrain the same attack-specific models on speaker-disjoint gender-composition splits with known speaker IDs and check whether the underrepresented-gender disadvantage and the fixed 1.317 pp EER gap under Oracle calibration still appear at the same magnitude.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper studies gender bias in audio deepfake detection on ASVspoof5 under a controlled custom 70/10/20 per-attack, per-gender re-split. It trains 384 attack-specific ResNet18 models (288 WavLM-Base+ across nine gender-composition configurations; 96 LogSpectrogram under three baselines) and evaluates six group-fairness metrics plus six post-hoc threshold calibrations, including an Oracle with evaluation labels. Main claims: (i) training gender composition strongly predicts bias direction—the underrepresented gender is worse at test time in 31/32 attacks; (ii) WavLM-Base+ yields EER gaps 3.0–4.3× larger than LogSpectrogram under matched training, and balanced training nearly removes LogSpec disparity but leaves WavLM largely biased; (iii) all six calibrations, including Oracle, leave mean ΔEER fixed at 1.317 pp, so distribution-level EER gaps require training-time fixes; (iv) attack type and metric choice change fairness rankings, so no single metric suffices.

Significance. If the results hold, this is a useful, large-scale controlled demonstration that gender fairness in spoofing countermeasures is primarily a data- and representation-level problem rather than a pure post-processing problem. Strengths include the systematic composition design (including the F+75%M vs M+25%F internal check), scale (384 models, 32 attacks), Poisson bootstrap with Benjamini–Hochberg correction, explicit multi-metric reporting, and practical deployment guidelines. The LogSpec vs WavLM contrast under identical composition is informative for the SSL-frontend literature. The work is incremental relative to the authors’ prior ASVspoof5 fairness papers [13], [24] but fills a clear gap by treating composition as the independent variable.

major comments (3)
  1. §III-F1–F3 and Table X / abstract: Per-gender EER is defined by finding, for each gender independently, the threshold where that gender’s FAR equals FRR. Under that definition, ΔEER is a property of the two score distributions (ROC shapes) and is invariant to any post-hoc choice of deployment thresholds, including Oracle. Reporting that all six strategies “leave the EER gap unchanged at 1.317 pp” is therefore largely definitional, not an empirical discovery. The substantive point—that thresholding cannot reshape group ROCs—is correct and worth making, but the abstract and §IV-C should reframe this as a definitional consequence of distribution-level EER and put empirical weight on the metric trade-offs (FPR gap, EOpD, SPD, etc.) and the composition→bias results, not on the numerical invariance of ΔEER under Oracle.
  2. Contribution 2 and §IV-B1 (Table VI): The claim that “combined training reduces the mean EER gap from 1.042 pp to 0.273 pp compared to within-attack balancing at a similar gender ratio” is not supported as stated. F+50%M (gap 1.042 pp) is not a similar gender ratio to Combined (100% F + 100% M). The nearer comparison is F+75%M (0.438 pp) or M+75%F (0.627 pp). Combined also uses more minority-gender data (and thus more total data) than any partial mix, so the extra drop to 0.273 pp confounds balance with sample size. The attribution to a special “population-level pooling” benefit beyond within-attack balancing should be revised or supported by a matched-N or matched-ratio control.
  3. §III-B and §V: Custom splits are stratified by attack and gender but not speaker-disjoint; evaluation speakers are expected to appear in training and overlap cannot be quantified. This does not overturn the directional composition→bias or the definitional ΔEER-invariance claims under fixed split construction, but it does limit interpretation of absolute EERs, cross-attack gaps, and any implication of speaker-independent fairness. Absolute gap magnitudes and cross-attack TED (already flagged as unstable) should be caveated more sharply in the abstract and conclusions so readers do not treat them as speaker-generalization results.
minor comments (6)
  1. §I and abstract: “we evaluated” should be “we evaluate” for tense consistency with the rest of the abstract.
  2. Table I / Fig. 1: Official ASVspoof5 partition labels (A01–A08 train, etc.) are shown alongside the custom re-split; a single sentence in the caption stating that official partitions are only the source pool would reduce confusion.
  3. §IV-B2: Cross-attack TED values (e.g., F: 0.115 → 54.785) are correctly called numerically unstable; consider moving them to an appendix or suppressing them in the main heatmap so they do not dominate Fig. 4.
  4. §III-G: N=1000 Poisson bootstrap resamples is a thin lower bound for claims at p≪0.001; stating this more prominently (as the authors already note for future work) would help readers calibrate confidence.
  5. Notation: SPD, EOpD, EOD, PPD, TED, FPRgap are defined clearly in §III-F2; adding a one-line “metric → deployment use-case” table early would help non-fairness readers navigate §IV-C1.
  6. References [13] and [24] are the authors’ own closely related work; ensure the novelty paragraph states explicitly what is new (composition as controlled IV; multi-composition calibration failure modes) versus reused metrics/protocol.

Circularity Check

0 steps flagged

Empirical composition-and-calibration study; no circular derivation chain

full rationale

The paper's load-bearing claims are experimental measurements: nine controlled gender-composition training sets, 384 attack-specific ResNet18 models (LogSpectrogram and WavLM-Base+), per-gender EER and six group-fairness metrics, and six post-hoc threshold strategies including Oracle. Bias direction tracking the underrepresented gender, the 3.0–4.3× larger WavLM gaps, and the invariance of ΔEER = 1.317 pp under every calibration (including Oracle) are read off the resulting score distributions and ROC curves; they are not obtained by fitting a parameter and re-labeling it as a prediction, nor by defining one quantity in terms of another. Self-citations [13] and [24] supply prior fixed-composition fairness numbers and the earlier balanced-training FPR-gap result that the present work extends; they are not used as uniqueness theorems or as the sole justification for the composition-variation or Oracle-EER findings. The known fact that a threshold cannot reshape a ROC (hence cannot change per-gender EER) is stated and then verified empirically, not smuggled in as a circular premise. Speaker non-disjointness of the custom splits is an external-validity limitation, not a circular step. No self-definitional, fitted-as-prediction, uniqueness-import, or ansatz-via-citation reductions appear.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

Empirical fairness study; load-bearing premises are dataset and protocol choices rather than free physical constants or new particles. Free parameters are standard training hyperparameters and the 0.5 pp practical-bias threshold. Axioms are domain conventions (binary ASVspoof5 gender labels, EER as primary disparity, attack-specific models). No invented entities.

free parameters (3)
  • learning_rate_and_weight_decay = 3e-5 / 1e-4
    AdamW lr=3e-5, weight decay=1e-4 chosen by authors; not fitted to fairness outcomes but affect absolute EERs.
  • practical_bias_threshold_0.5pp = 0.5 pp
    Authors treat EER gap >0.5 pp as practically meaningful bias when ranking configurations; this cutoff is conventional, not derived.
  • bootstrap_N_and_alpha = N=1000, alpha=0.05
    N=1000 Poisson resamples and BH alpha=0.05 set significance claims; acknowledged as a lower bound for p<0.001.
axioms (4)
  • domain assumption Official ASVspoof5 binary gender labels correctly partition speakers into female and male for fairness evaluation.
    Invoked throughout Section III-A/B and V; non-binary and intersectional identities are excluded by protocol.
  • domain assumption Absolute EER gap is a valid primary indicator of distribution-level gender disparity that threshold calibration cannot change.
    Stated in III-F and used as the invariant quantity in Table X and the conclusion.
  • ad hoc to paper Attack-specific models isolate attack-dependent bias without requiring speaker-independent generalization.
    Section III-E and V explicitly prioritize composition control over speaker-disjoint evaluation.
  • standard math Poisson bootstrap on confusion-matrix cells adequately approximates utterance-level uncertainty for the reported p-values.
    Section III-G cites Hanley & MacGibbon; validity assumed for the evaluation set sizes used.

pith-pipeline@v1.1.0-grok45 · 29327 in / 2803 out tokens · 30667 ms · 2026-07-14T14:47:00.285921+00:00 · methodology

0 comments
read the original abstract

Audio deepfake detection models determine whether speech is genuine or artificially generated, but high overall accuracy can mask substantial performance disparities across demographic groups. In this work, we investigate gender bias in audio deepfake detection using the ASVspoof5 dataset. We use ASVspoof5 under a controlled custom split designed to isolate gender-composition effects. We train attack-specific models on nine training sets with different gender compositions, ranging from female-only to male-only. We use a ResNet18 classifier with LogSpectrogram and WavLM-Base+ features, and we evaluated six post-hoc threshold calibration methods. Experimental results show that training data composition strongly predicts bias direction, with the underrepresented gender performing worse at test time. WavLM-Base+ features are shown to produce gender performance gaps 3.0 to 4.3 times larger than LogSpectrogram under identical training conditions, and balanced training is found to reduce LogSpectrogram bias but leave WavLM bias largely intact. Moreover, all six calibration strategies, including Oracle calibration with full test-set label access, leave the Equal Error Rate gap unchanged at 1.317 pp, confirming that threshold adjustment cannot correct underlying score distribution disparities. Overall, these findings suggest that gender fairness in audio deepfake detection must be addressed at training time, as post-hoc methods can only partially mitigate the resulting disparities

Figures

Figures reproduced from arXiv: 2607.09891 by Aishwarya R. Fursule, Anderson R. Avila, Shruti Kshirsagar, Vamshi Nallaguntla.

Figure 1
Figure 1. Figure 1: ASVspoof5 dataset protocol used in this study. (A01-A08 training, A09 -A16 development, A17-A32 evaluation). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed system for gender fairness evaluation and post-hoc mitigation in audio deepfake detection. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: EERgap under same-attack and cross-attack evaluation across all nine gender-composition configurations for TTS (text-to-speech), VC (voice conversion), and AT (adversarial perturbation) attacks. tics [44], a condition that holds when tests share overlapping utterance pools as in this study. IV. RESULTS & DISCUSSION In this section, we present and discuss our experimental results in three parts: feature com… view at source ↗
Figure 4
Figure 4. Figure 4: Fairness metric heatmap across TTS (text-to-speech), VC (voice conversion), and AT (adversarial perturbation) attacks under same-attack (left side) [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 14 linked inside Pith

  1. [1]

    Warning: Humans cannot reliably detect speech deepfakes,

    K. T. Mai, S. Bray, T. Davies, and L. D. Griffin, “Warning: Humans cannot reliably detect speech deepfakes,”PLOS ONE, vol. 18, no. 8, p. e0285333, 2023

    2023

  2. [2]

    Audio deepfake detection using deep learning,

    O. A. Shaaban and R. Yildirim, “Audio deepfake detection using deep learning,”Engineering Reports, vol. 7, no. 3, p. e70087, 2025

    2025

  3. [3]

    A comprehensive survey of deepfake generation and detection techniques in audio-visual media,

    I. Khan, K. Khan, and A. Ahmad, “A comprehensive survey of deepfake generation and detection techniques in audio-visual media,”ICCK Journal of Image Analysis and Processing, vol. 1, no. 2, pp. 73–95, 2025

    2025

  4. [4]

    Audio deepfake detection: A survey,

    J. Yi, C. Wang, J. Tao, X. Zhang, C. Y . Zhang, and Y . Zhao, “Audio deepfake detection: A survey,”arXiv preprint arXiv:2308.14970, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Vulnerabilities of audio-based biometric authentication systems against deepfake speech synthesis,

    M. Hong, D. Jiang, Z. Xie, W. Zhao, G. Wang, and C. J. Zhang, “Vulnerabilities of audio-based biometric authentication systems against deepfake speech synthesis,”arXiv preprint arXiv:2601.02914, 2026

    arXiv 2026

  6. [6]

    Audio deepfake detection: What has been achieved and what lies ahead,

    B. Zhang, H. Cui, V . Nguyen, and M. Whitty, “Audio deepfake detection: What has been achieved and what lies ahead,”Sensors, vol. 25, no. 7, p. 1989, 2025

    1989

  7. [7]

    ASVspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,

    X. Wang, H. Delgado, H. Tak, J. W. Jung, H. J. Shim, M. Todisco, et al., “ASVspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,”arXiv preprint arXiv:2408.08739, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    End-to-end anti-spoofing with RawNet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with RawNet2,” inProc. ICASSP 2021 – IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6369–6373

    2021

  9. [9]

    AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

    J. W. Jung, H. S. Heo, H. Tak, H. J. Shim, J. S. Chung, B. J. Lee, et al., “AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” inProc. ICASSP 2022 – IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6367–6371

    2022

  10. [10]

    Bias in data-driven artificial intelligence systems: An introductory survey,

    E. Ntoutsi, P. Fafalios, U. Gadiraju, V . Iosifidis, W. Nejdl, M. E. Vidal, et al., “Bias in data-driven artificial intelligence systems: An introductory survey,”Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 10, no. 3, p. e1356, 2020

    2020

  11. [11]

    Fair voice biometrics: Impact of demographic imbalance on group fairness in speaker recogni- tion,

    G. Fenu, M. Marras, G. Medda, and G. Meloni, “Fair voice biometrics: Impact of demographic imbalance on group fairness in speaker recogni- tion,” inProc. Interspeech, 2021, pp. 1892–1896

    2021

  12. [12]

    Real-time detection of AI-generated speech for deepfake voice conversion,

    J. J. Bird and A. Lotfi, “Real-time detection of AI-generated speech for deepfake voice conversion,”arXiv preprint arXiv:2308.12734, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Gender fairness in audio deepfake detection: Performance and disparity analysis,

    A. Fursule, S. Kshirsagar, and A. R. Avila, “Gender fairness in audio deepfake detection: Performance and disparity analysis,” inProc. 2026 IEEE Conference on Artificial Intelligence (CAI), 2026, pp. 2116–2121

    2026

  14. [14]

    Fairness without demographics in repeated loss minimization,

    T. Hashimoto, M. Srivastava, H. Namkoong, and P. Liang, “Fairness without demographics in repeated loss minimization,” inProc. Interna- tional Conference on Machine Learning, 2018, pp. 1929–1938

    2018

  15. [15]

    AFSS: Artifact-focused self-synthesis for mitigat- ing bias in audio deepfake detection,

    H. S. Nguyen-Le, H. C. Nguyen-Thanh, N. A. Le-Khac, D. T. Nguyen, and H. H. Nguyen-Le, “AFSS: Artifact-focused self-synthesis for mitigat- ing bias in audio deepfake detection,”arXiv preprint arXiv:2603.26856, 2026

    arXiv 2026

  16. [16]

    GBDF: Gender balanced deepfake dataset towards fair deepfake detection,

    A. V . Nadimpalli and A. Rattani, “GBDF: Gender balanced deepfake dataset towards fair deepfake detection,” inProc. International Confer- ence on Pattern Recognition, 2022, pp. 320–337

    2022

  17. [17]

    A survey on bias and fairness in machine learning,

    N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan, “A survey on bias and fairness in machine learning,”ACM Computing Surveys, vol. 54, no. 6, pp. 1–35, 2021

    2021

  18. [18]

    Improving fairness in deepfake detection,

    Y . Ju, S. Hu, S. Jia, G. H. Chen, and S. Lyu, “Improving fairness in deepfake detection,” inProc. IEEE/CVF Winter Conference on Applica- tions of Computer Vision, 2024, pp. 4655–4665

    2024

  19. [19]

    Preserving fairness generalization in deepfake detection,

    L. Lin, X. He, Y . Ju, X. Wang, F. Ding, and S. Hu, “Preserving fairness generalization in deepfake detection,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16815–16825

    2024

  20. [20]

    FairSSD: Understanding bias in synthetic speech detectors,

    A. K. S. Yadav, K. Bhagtani, D. Salvi, P. Bestagini, and E. J. Delp, “FairSSD: Understanding bias in synthetic speech detectors,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 4418–4428

    2024

  21. [21]

    Equality of opportunity in supervised learning,

    M. Hardt, E. Price, and N. Srebro, “Equality of opportunity in supervised learning,” inAdvances in Neural Information Processing Systems, vol. 29, 2016

    2016

  22. [22]

    Fair prediction with disparate impact: A study of bias in recidivism prediction instruments,

    A. Chouldechova, “Fair prediction with disparate impact: A study of bias in recidivism prediction instruments,”Big Data, vol. 5, no. 2, pp. 153– 163, 2017

    2017

  23. [23]

    Phonetic analysis of real and synthetic speech using HuBERT embeddings: Perspectives for deepfake detection,

    D. E. Temmar, A. Hamadene, V . Nallaguntla, A. Fursule, M. S. Allili, S. Kshirsagar, and A. R. Avila, “Phonetic analysis of real and synthetic speech using HuBERT embeddings: Perspectives for deepfake detection,” inProc. 2025 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2025, pp. 86–91

    2025

  24. [24]

    Towards trustworthy audio deepfake detection: A systematic framework for diagnosing and mitigating gender bias,

    A. Fursule, S. Kshirsagar, and A. R. Avila, “Towards trustworthy audio deepfake detection: A systematic framework for diagnosing and mitigating gender bias,” inProc. IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2026

    2026

  25. [25]

    PhonemeDF: A synthetic speech dataset for audio deepfake detection and naturalness evaluation,

    V . Nallaguntla, A. Fursule, S. Kshirsagar, and A. R. Avila, “PhonemeDF: A synthetic speech dataset for audio deepfake detection and naturalness evaluation,”arXiv preprint arXiv:2603.15037, 2026

    arXiv 2026

  26. [26]

    Investigating the impact of speech enhancement on audio deepfake detection in noisy environments,

    S. Kshirsagar and A. R. Avila, “Investigating the impact of speech enhancement on audio deepfake detection in noisy environments,”arXiv preprint arXiv:2603.14767, 2026

    arXiv 2026

  27. [27]

    An examination of fairness of AI models for deepfake detection,

    L. Trinh and Y . Liu, “An examination of fairness of AI models for deepfake detection,”arXiv preprint arXiv:2105.00558, 2021

    Pith/arXiv arXiv 2021

  28. [28]

    Analyzing fairness in deepfake detection with massively annotated databases,

    Y . Xu, P. Terh ¨orst, M. Pedersen, and K. Raja, “Analyzing fairness in deepfake detection with massively annotated databases,”IEEE Transac- tions on Technology and Society, vol. 5, no. 1, pp. 93–106, 2024

    2024

  29. [29]

    Bias in automated speaker recognition,

    W. T. Hutiri and A. Y . Ding, “Bias in automated speaker recognition,” inProc. ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2022, pp. 230–247

    2022

  30. [30]

    SCDF: A Speaker Char- acteristics DeepFake Speech Dataset for Bias Analysis,

    V . Stan ˇek, K. Srna, A. Firc, and K. Malinka, “SCDF: A Speaker Char- acteristics DeepFake Speech Dataset for Bias Analysis,”arXiv preprint arXiv:2508.07944, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  32. [32]

    Easy, interpretable, effective: openSMILE for voice deepfake detection,

    O. Pascu, D. Oneat ¸ ˘a, H. Cucu, and N. M ¨uller, “Easy, interpretable, effective: openSMILE for voice deepfake detection,” inProc. ICASSP 2025 – IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  33. [33]

    Natural- Speech: End-to-end text-to-speech synthesis with human-level quality,

    X. Tan, J. Chen, H. Liu, J. Cong, C. Zhang, Y . Liu, et al., “Natural- Speech: End-to-end text-to-speech synthesis with human-level quality,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 6, pp. 4234–4245, 2024

    2024

  34. [34]

    ASVspoof: The automatic speaker verification spoofing and countermeasures challenge,

    Z. Wu, J. Yamagishi, T. Kinnunen, C. Hanilc ¸i, M. Sahidullah, A. Sizov, et al., “ASVspoof: The automatic speaker verification spoofing and countermeasures challenge,”IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 4, pp. 588–604, 2017

    2017

  35. [35]

    ASVspoof 2021: Accelerating progress in spoofed and deepfake speech detection,

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, et al., “ASVspoof 2021: Accelerating progress in spoofed and deepfake speech detection,”arXiv preprint arXiv:2109.00537, 2021

    Pith/arXiv arXiv 2021

  36. [36]

    ASVspoof 2021: Towards spoofed and deepfake speech detection in the wild,

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kinnunen, et al., “ASVspoof 2021: Towards spoofed and deepfake speech detection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2507–2522, 2023

    2021

  37. [37]

    To train or not to train adversarially: A study of bias mitigation strategies for speaker recognition,

    R. Peri, K. Somandepalli, and S. Narayanan, “To train or not to train adversarially: A study of bias mitigation strategies for speaker recognition,”arXiv preprint arXiv:2203.09122, 2022

    Pith/arXiv arXiv 2022

  38. [38]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  39. [39]

    Controlling the false discovery rate: A practical and powerful approach to multiple testing,

    Y . Benjamini and Y . Hochberg, “Controlling the false discovery rate: A practical and powerful approach to multiple testing,”Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 1, pp. 289–300, 1995

    1995

  40. [40]

    An intervention-based framework for shortcut diagnosis in spoofing countermeasures,

    S. Rubio, P. Bello, D. Ribas, A. Miguel, E. Lleida, and A. Ortega, “An intervention-based framework for shortcut diagnosis in spoofing countermeasures,” inProc. Odyssey 2026: The Speaker and Language Recognition Workshop, 2026, pp. 1–8

    2026

  41. [41]

    Can SSL frontend generalize to all-type audio spoofing?

    A. Das, Y . El Kheir, F. R. Guttierez, T. Polzehl, and S. M¨oller, “Can SSL frontend generalize to all-type audio spoofing?” inProc. Odyssey 2026: The Speaker and Language Recognition Workshop, 2026, pp. 277–283

    2026

  42. [42]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  43. [43]

    Creating non-parametric bootstrap samples using Poisson frequencies,

    J. A. Hanley and B. MacGibbon, “Creating non-parametric bootstrap samples using Poisson frequencies,”Computer Methods and Programs in Biomedicine, vol. 83, no. 1, pp. 57–62, 2006

    2006

  44. [44]

    The control of the false discovery rate in multiple testing under dependency,

    Y . Benjamini and D. Yekutieli, “The control of the false discovery rate in multiple testing under dependency,”The Annals of Statistics, vol. 29, no. 4, pp. 1165–1188, 2001. 15

    2001

  45. [45]

    Efron and R

    B. Efron and R. J. Tibshirani,An Introduction to the Bootstrap. New York, NY , USA: Chapman and Hall, 1993

    1993

  46. [46]

    Creating new language and voice components for the updated MaryTTS text-to-speech synthesis platform,

    I. Steiner and S. Le Maguer, “Creating new language and voice components for the updated MaryTTS text-to-speech synthesis platform,” inProceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan, May 2018

    2018

  47. [47]

    ZMM-TTS: Zero-shot multilingual and multispeaker speech synthesis conditioned on self-supervised discrete speech representations,

    C. Gong, X. Wang, E. Cooper, D. Wells, L. Wang, J. Dang, and J. Ya- magishi, “ZMM-TTS: Zero-shot multilingual and multispeaker speech synthesis conditioned on self-supervised discrete speech representations,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 4036–4051, 2024

    2024

  48. [48]

    YourTTS: Towards zero-shot multi-speaker TTS and zero- shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “YourTTS: Towards zero-shot multi-speaker TTS and zero- shot voice conversion for everyone,” inProceedings of the International Conference on Machine Learning (ICML), pp. 2709–2720, PMLR, Jun. 2022

    2022

  49. [49]

    XTTS: A massively multilingual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, and J. Weber, “XTTS: A massively multilingual zero-shot text-to-speech model,”arXiv preprint arXiv:2406.04904, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Glow-TTS: A generative flow for text-to-speech via monotonic alignment search,

    J. Kim, S. Kim, J. Kong, and S. Yoon, “Glow-TTS: A generative flow for text-to-speech via monotonic alignment search,” inAdvances in Neural Information Processing Systems, vol. 33, pp. 8067–8077, 2020

    2020

  51. [51]

    Grad- TTS: A diffusion probabilistic model for text-to-speech,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, and M. Kudinov, “Grad- TTS: A diffusion probabilistic model for text-to-speech,” inProceedings of the International Conference on Machine Learning (ICML), pp. 8599– 8608, PMLR, Jul. 2021

    2021

  52. [52]

    BigVGAN: A universal neural vocoder with large-scale training,

    S. G. Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “BigVGAN: A universal neural vocoder with large-scale training,”arXiv preprint arXiv:2206.04658, 2022

    Pith/arXiv arXiv 2022

  53. [53]

    Exact prosody cloning in zero-shot multispeaker text-to-speech,

    F. Lux, J. Koch, and N. T. Vu, “Exact prosody cloning in zero-shot multispeaker text-to-speech,” inProceedings of the 2022 IEEE Spoken Language Technology Workshop (SLT), pp. 962–969, IEEE, Jan. 2023

    2022

  54. [54]

    FastPitch: Parallel text-to-speech with pitch prediction,

    A. Ła ´ncucki, “FastPitch: Parallel text-to-speech with pitch prediction,” in Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6588–6592, IEEE, Jun. 2021

    2021

  55. [55]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” inProceedings of the International Conference on Machine Learning (ICML), pp. 5530–5540, PMLR, Jul. 2021

    2021

  56. [56]

    Low-resource multilingual and zero- shot multispeaker TTS,

    F. Lux, J. Koch, and N. T. Vu, “Low-resource multilingual and zero- shot multispeaker TTS,” inProceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 741–751, Nov. 2022

    2022

  57. [57]

    Diffusion-based voice conversion with fast maximum likelihood sam- pling scheme,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, M. Kudinov, and J. Wei, “Diffusion-based voice conversion with fast maximum likelihood sam- pling scheme,”arXiv preprint arXiv:2109.13821, 2021

    Pith/arXiv arXiv 2021

  58. [58]

    HiFi-GAN: Generative adversarial net- works for efficient and high-fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative adversarial net- works for efficient and high-fidelity speech synthesis,” inAdvances in Neural Information Processing Systems, vol. 33, pp. 17022–17033, 2020

    2020

  59. [59]

    Natural TTS synthesis by conditioning WaveNet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerry-Ryan, R. A. Saurous, Y . Agiomyrgiannakis, and Y . Wu, “Natural TTS synthesis by conditioning WaveNet on mel spectrogram predictions,” inProceedings of the IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4779– 47...

    2018

  60. [60]

    StarGANv2-VC: A diverse, unsu- pervised, non-parallel framework for natural-sounding voice conversion,

    Y . A. Li, A. Zare, and N. Mesgarani, “StarGANv2-VC: A diverse, unsu- pervised, non-parallel framework for natural-sounding voice conversion,” arXiv preprint arXiv:2107.10394, 2021

    Pith/arXiv arXiv 2021

  61. [61]

    V oice conversion using speech-to-speech neuro-style transfer,

    E. A. AlBadawy and S. Lyu, “V oice conversion using speech-to-speech neuro-style transfer,” inProc. Interspeech, 2020, pp. 4726–4730

    2020

  62. [62]

    Self-supervised speech representation learning: A review,

    A. Mohamed, H.-Y . Lee, L. Borgholt, J. D. Havtorn, J. Edin, C. Igel, K. Kirchhoff, S.-W. Li, K. Livescu, L. Mangu, T. N. Sainath, and S. Watanabe, “Self-supervised speech representation learning: A review,” IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1179–1210, 2022

    2022

  63. [63]

    Toward noise-aware audio deepfake detection: Survey, SNR-benchmarks, and practical recipes,

    U. Sen, A. Luqman, and A. Chattopadhyay, “Toward noise-aware audio deepfake detection: Survey, SNR-benchmarks, and practical recipes,” arXiv preprint arXiv:2512.13744, 2025

    arXiv 2025

  64. [64]

    C3-DINO: Joint contrastive and non-contrastive self-supervised learning for speaker verification,

    C. Zhang and D. Yu, “C3-DINO: Joint contrastive and non-contrastive self-supervised learning for speaker verification,”IEEE Journal of Se- lected Topics in Signal Processing, vol. 16, no. 6, pp. 1273–1283, 2022

    2022

  65. [65]

    Context and transcripts improve detection of deepfake audios of public figures,

    C. Gao, M. Postiglione, J. Baldwin, N. Denisenko, I. Gortner, L. Fosdick, and V . S. Subrahmanian, “Context and transcripts improve detection of deepfake audios of public figures,”arXiv preprint arXiv:2601.13464, 2026

    arXiv 2026

  66. [66]

    Fine-tuning self-supervised learning models for end-to-end pronunciation scoring,

    A. I. Zahran, A. A. Fahmy, K. T. Wassif, and H. Bayomi, “Fine-tuning self-supervised learning models for end-to-end pronunciation scoring,” IEEE Access, vol. 11, pp. 112650–112663, 2023

    2023

  67. [67]

    Using optimal f-measure and random resampling in gene ontology enrichment calculations,

    W. Ge, Z. Fazal, and E. Jakobsson, “Using optimal f-measure and random resampling in gene ontology enrichment calculations,”Frontiers in Applied Mathematics and Statistics, vol. 5, p. 20, 2019

    2019

  68. [68]

    Modified FDR controlling proce- dure for multi-stage analyses,

    C. Tuglus and M. J. van der Laan, “Modified FDR controlling proce- dure for multi-stage analyses,”Statistical Applications in Genetics and Molecular Biology, vol. 8, no. 1, Art. 12, 2009

    2009

  69. [69]

    Cyclostationarity analysis as a complement to self-supervised representations for speech deepfake detection,

    C. Hanilc ¸i, M. Sahidullah, and T. Kinnunen, “Cyclostationarity analysis as a complement to self-supervised representations for speech deepfake detection,”arXiv preprint arXiv:2603.03921, 2026

    arXiv 2026

  70. [70]

    Training-free cross- lingual dysarthria severity assessment via phonological subspace analysis in self-supervised speech representations,

    B. Muller, A. A. Ortiz Barra ˜n´on, and L. Roberts, “Training-free cross- lingual dysarthria severity assessment via phonological subspace analysis in self-supervised speech representations,”medRxiv, 2026

    2026

  71. [71]

    Bootstrap confidence regions for the intensity of a Poisson point process,

    A. Cowling, P. Hall, and M. J. Phillips, “Bootstrap confidence regions for the intensity of a Poisson point process,”Journal of the American Statistical Association, vol. 91, no. 436, pp. 1516–1524, 1996

    1996

  72. [72]

    PyTorch: An imperative style, high- performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K ¨opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “PyTorch: An imperative style, high- performance deep learning library,” inAdvances in Neural Information Processing...

    2019

  73. [73]

    Fairness through awareness,

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, and R. Zemel, “Fairness through awareness,” inProceedings of the 3rd Innovations in Theoretical Computer Science Conference, 2012, pp. 214–226

    2012

  74. [74]

    Phoneme-Level Deep- fake Detection Across Emotional Conditions Using Self-Supervised Em- beddings,

    V . Nallaguntla, S. Kshirsagar, and A. R. Avila, “Phoneme-Level Deep- fake Detection Across Emotional Conditions Using Self-Supervised Em- beddings,”arXiv preprint arXiv:2605.03079, 2026

    Pith/arXiv arXiv 2026

  75. [75]

    A review on fairness in machine learning,

    D. Pessach and E. Shmueli, “A review on fairness in machine learning,” ACM Computing Surveys, vol. 55, no. 3, pp. 1–44, 2022

    2022

  76. [76]

    Fairness definitions explained,

    S. Verma and J. Rubin, “Fairness definitions explained,” inProceedings of the International Workshop on Software Fairness, pp. 1–7, May 2018

    2018

  77. [77]

    Measuring algorithmic fairness,

    D. Hellman, “Measuring algorithmic fairness,”Virginia Law Review, vol. 106, no. 4, pp. 811–866, 2020

    2020

  78. [78]

    Bias preservation in machine learning: The legality of fairness metrics under EU non-discrimination law,

    S. Wachter, B. Mittelstadt, and C. Russell, “Bias preservation in machine learning: The legality of fairness metrics under EU non-discrimination law,”West Virginia Law Review, vol. 123, no. 3, pp. 735–790, 2021

    2021

  79. [79]

    Inclusive speaker verification with adaptive thresholding,

    N. Jain and H. Wang, “Inclusive speaker verification with adaptive thresholding,”arXiv preprint arXiv:2111.05501, 2021

    Pith/arXiv arXiv 2021

  80. [80]

    On fairness and calibration,

    G. Pleiss, M. Raghavan, F. Wu, J. Kleinberg, and K. Q. Weinberger, “On fairness and calibration,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017