Pith. sign in

REVIEW 2 major objections 5 minor 53 references

Deepfake detectors lean on specific vowels, fricatives, and pauses in ways that change with the attack and the speaker.

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

2026-07-10 04:46 UTC pith:2F4FZHXL

load-bearing objection Solid methodological pipeline that turns Grad-CAM into phoneme-level stats on ASVspoof 5; the findings are useful and carefully reported, but remain correlational rather than causal. the 2 major comments →

arxiv 2607.08586 v1 pith:2F4FZHXL submitted 2026-07-09 eess.AS cs.SD

Why Do You Say It Like That? A Phoneme-Level Framework for Explainable Speech Deepfake Detection

Anna Taylor , Michele Panariello , Massimiliano Todisco , Chiara Galdi , Nicholas Evans , Driss Matrouf This is my paper
classification eess.AS cs.SD
keywords speech deepfake detectionexplainable AIphoneme-level attributionGrad-CAMASVspoof 5self-supervised speech representationsforced alignment
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.

Modern speech deepfake detectors that use self-supervised features achieve strong accuracy, yet they rarely explain their decisions in terms a human can follow. This paper supplies a post-hoc pipeline that maps Gradient-weighted Class Activation Mapping saliency onto phoneme and pause boundaries obtained by automatic transcription and forced alignment. Applied to a competitive WavLM-plus-CNN detector on the ASVspoof 5 evaluation set, the method shows that support for spoof versus bona-fide decisions is not uniform: vowels such as /oU/, /E/ and /AI/, fricatives such as /s/ and /f/, and non-speech regions carry the largest attack-dependent and speaker-dependent attribution mass. The same framework also reveals that bona-fide evidence concentrates near phoneme boundaries while spoof evidence concentrates inside steady-state segments. Detection performance remains competitive (pooled EER 8.52 percent), so the linguistic interpretations do not come at the cost of accuracy. A sympathetic reader cares because the resulting explanations are stated in ordinary phonetic units rather than opaque spectrogram patches, making detector behavior inspectable across speakers and generation systems.

Core claim

A Grad-CAM-plus-forced-alignment pipeline can convert the temporal saliency of a convolutional deepfake detector into statistically significant, frequency-normalized phoneme-level importance scores. On ASVspoof 5 these scores vary systematically by attack and by speaker, with the largest effects concentrated on particular vowels, fricatives and non-speech regions, while the detector itself retains competitive equal-error-rate performance.

What carries the argument

Phoneme-aligned Grad-CAM: class-specific saliency maps from the final convolutional layer are averaged inside forced-alignment phoneme and silence spans, then aggregated and tested with Kruskal-Wallis and effect-size statistics to yield human-readable phonetic importance scores.

Load-bearing premise

Averaging Grad-CAM values inside forced-alignment phoneme boundaries is assumed to identify the linguistic units the detector actually relies on, rather than merely correlated acoustic properties such as duration or energy.

What would settle it

Replace the forced-alignment phoneme spans with random or deliberately misaligned boundaries of identical length; if the attack- and speaker-dependent effect sizes collapse, the claimed phoneme-level attributions are artifacts of the alignment rather than genuine detector cues.

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

2 major / 5 minor

Summary. The paper proposes a post-hoc phoneme-level explainability pipeline for CNN-based speech deepfake detectors. An SSL front-end (WavLM Base+) feeds a temporal CNN classifier; Grad-CAM is applied separately to the bona-fide and spoof logits, and the resulting temporal saliency is averaged inside phoneme (and silence) spans obtained from Whisper transcription plus Bournemouth Forced Aligner. Aggregated scores are subjected to Kruskal–Wallis tests with Benjamini–Hochberg correction and epsilon-squared effect sizes, revealing attack- and speaker-dependent patterns (strongest for vowels /oU/, /E/, /AI/ and fricatives /s/, /f/, ε² up to 0.135) while the detector itself reaches a competitive pooled EER of 8.52 % on ASVspoof 5 evaluation data. The authors also examine temporal localization of peaks, non-speech mass, and gender/speaker effects, and they explicitly note that Grad-CAM remains correlational.

Significance. If the mapping from Grad-CAM averages to linguistically meaningful units is accepted as useful (even if not strictly causal), the work supplies a practical, architecture-agnostic bridge between modern SSL deepfake detectors and human-interpretable phonetic categories. The statistical apparatus (frequency normalization, non-parametric tests, effect-size reporting) is carefully applied at scale on ASVspoof 5, and the observation that different attacks and speakers concentrate evidence on different phonemes and on boundary versus steady-state regions is of genuine interest to both the anti-spoofing and forensic communities. The framework does not claim a new state-of-the-art detector; its value lies in the interpretability layer that can be attached to existing CNN systems without performance loss.

major comments (2)
  1. Section VI and the central interpretability claim: every phoneme-importance result (Table I, Fig. 3, non-speech mass percentages) rests on the premise that averaging Grad-CAM values inside forced-alignment spans isolates the linguistic units the detector relies on. The authors themselves state that Grad-CAM “does not establish direct causal attribution and it is possible that highlighted phonemes may reflect correlated acoustic properties.” Frequency normalization removes only occurrence-rate confounds, not duration, energy, pitch or coarticulation confounds. Without at least one controlled ablation (e.g., duration-matched resampling, pitch-normalized resynthesis, or a causal intervention that alters only the named phoneme while holding low-level acoustics fixed) the Kruskal–Wallis / ε² findings remain correlational. This is the single load-bearing soft spot for the claim that the pipeli
  2. Section III-C / IV-A: forced-alignment quality on strongly synthetic speech is not quantified. Whisper + Bournemouth Forced Aligner errors are acknowledged as a limitation, yet no phone-error-rate or boundary-offset statistics are reported for the 16 ASVspoof 5 attack conditions. Because the most challenging attacks (A24, A28) already show EERs >20 %, systematic misalignment on those conditions could inflate or deflate the very phonemes declared most discriminative. A short validation (manual inspection of a stratified subset or comparison against an oracle alignment) is needed before the attack-dependent rankings can be trusted.
minor comments (5)
  1. Table I caption and surrounding text: “frequency-normalized” is defined only in prose; a one-line formula (e.g., score_p / count_p) would make the procedure fully reproducible.
  2. Figure 2: the claim that boundary versus interior peaks “represent over 2 % of the CAM peaks” is unclear; the figure itself does not show percentages and the text does not define the denominator.
  3. Section V-A: per-attack EERs are summarized qualitatively (“most <8 %, A24/A28 >20 %”) but never tabulated; a compact table would strengthen the link between detection difficulty and the phoneme-level patterns.
  4. Terminology note in the introduction (phoneme vs. phone) is welcome, yet the remainder of the paper continues to use “phoneme” for acoustic realizations; a consistent parenthetical reminder would help non-linguist readers.
  5. References [40] and [45] both cite Benjamini–Hochberg 1995; one entry is redundant.

Circularity Check

0 steps flagged

No circularity: post-hoc Grad-CAM attributions on held-out data are aggregated and tested statistically; nothing reduces a claimed prediction to a fitted input or self-definition by construction.

full rationale

The paper trains a frozen-WavLM + temporal-CNN detector on the ASVspoof 5 training partition, then extracts bona-fide-CAM and spoof-CAM maps on the speaker-disjoint evaluation set. Phoneme spans come from an independent Whisper transcript + Bournemouth Forced Aligner; importance scores are simple averages of the CAM values inside those spans. Kruskal-Wallis / epsilon-squared tests and chi-square tests are then applied to the resulting score distributions across attacks and speakers. No free parameter is fitted to the phoneme-importance statistics themselves, no uniqueness theorem is imported from the authors’ prior work to force the result, and no equation equates a “prediction” to an input by definition. Self-citations (ASVspoof 5, Malafide) supply only the corpus and attack labels; they are not load-bearing for the explanatory claims. The authors themselves note that Grad-CAM is correlational, which is a correctness caveat, not circularity. The derivation chain is therefore self-contained and observational.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The paper is methodological and empirical rather than theoretical. Its central claims rest on the validity of two external tools (Grad-CAM and forced alignment) when applied to synthetic speech, plus standard non-parametric statistics. No new physical entities or free parameters are fitted to produce the reported phonetic rankings; the detector’s own training hyper-parameters affect absolute EER but not the relative attribution patterns that form the main claim.

free parameters (1)
  • CNN channel count and training schedule (512 channels, 200 epochs Adam cosine)
    Architectural and optimization choices that determine the detector whose activations are later explained; not fitted to the phoneme-importance statistics themselves, yet they condition every attribution map.
axioms (3)
  • domain assumption Grad-CAM applied to the final convolutional layer yields temporally meaningful class-specific importance that can be averaged inside phoneme spans
    Invoked throughout Section III-B/C and all of Section V; known limitations (correlation vs causation, architecture sensitivity) are acknowledged only later in Section VI.
  • domain assumption Whisper transcription + Bournemouth Forced Aligner produce sufficiently accurate phoneme and silence boundaries even for strongly synthetic or adversarial speech
    Load-bearing for every phoneme-level score; errors would systematically bias the reported attack- and speaker-dependent patterns (Section III-C).
  • standard math Kruskal-Wallis tests with Benjamini-Hochberg correction and epsilon-squared effect sizes are appropriate for comparing attribution-score distributions across attacks
    Used for all significance claims in Section V-B and subsequent subsections.

pith-pipeline@v1.1.0-grok45 · 17544 in / 2953 out tokens · 43424 ms · 2026-07-10T04:46:01.079676+00:00 · methodology

0 comments
read the original abstract

As the accuracy of speech deepfake detection improves with the use of self-supervised representations such as wav2vec 2.0 and HuBERT, understanding why the speech is classified as bona fide or deepfake remains an open challenge. In pursuit of more trustworthy and interpretable artificial intelligence, we introduce a phoneme-level analysis framework that connects model predictions to measurable phonetic units. Our post-hoc explainability method is generally applicable to a variety of speech deepfake detection systems based on convolutional neural networks since it leverages Gradient-weighted Class Activation Mapping in conjunction with speech recognition to generate saliency maps aligned with phonemes and pauses. This pipeline reveals statistically significant attack- and speaker-dependent phonetic cues associated with spoofed speech in terms that humans can understand. Experiments using ASVspoof 5 show comparable detection performance to similar architectures while providing linguistic interpretations across speakers and spoofing conditions.

Figures

Figures reproduced from arXiv: 2607.08586 by Anna Taylor, Chiara Galdi, Driss Matrouf, Massimiliano Todisco, Michele Panariello, Nicholas Evans.

Figure 1
Figure 1. Figure 1: Example of phoneme-aligned Grad-CAM attribution maps for the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Grad-CAM activation peaks differ in their relative location within [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example of the most discriminative phonemes for three representative [PITH_FULL_IMAGE:figures/full_fig_p006_3.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

53 extracted references · 53 canonical work pages · 5 internal anchors

  1. [1]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, pp. 12 449–12 460

    work page 2020

  2. [2]

    HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 29, p. 3451–3460, Oct. 2021

    work page 2021

  3. [3]

    WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, M. Zeng, and F. Wei, “WavLM: Large-Scale Self-Supervised Pre- Training for Full Stack Speech Processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, pp. 1505–1518, 2021. [Online]. Available: https://api....

    work page 2021

  4. [4]

    The mythos of model interpretability,

    Z. C. Lipton, “The mythos of model interpretability,”Commun. ACM, vol. 61, no. 10, p. 36–43, Sep. 2018. [Online]. Available: https://doi.org/10.1145/3233231

    work page doi:10.1145/3233231 2018

  5. [5]

    On the Use/Misuse of the Term 'Phoneme'

    R. K. Moore and L. Skidmore, “On the use/misuse of the term ’phoneme’,”CoRR, vol. abs/1907.11640, 2019. [Online]. Available: http://arxiv.org/abs/1907.11640

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  6. [6]

    Phoneme duration modeling using speech rhythm-based speaker embeddings for multi-speaker speech synthesis,

    K. Fujita, A. Ando, and Y . Ijima, “Phoneme duration modeling using speech rhythm-based speaker embeddings for multi-speaker speech synthesis,” inProceedings of Interspeech 2021. ISCA, 2021, pp. 3141– 3145

    work page 2021

  7. [7]

    The Effect of Silence and Dual- Band Fusion in Anti-Spoofing System,

    Y . Zhang, W. Wang, and P. Zhang, “The Effect of Silence and Dual- Band Fusion in Anti-Spoofing System,” inInterspeech 2021, 2021, pp. 4279–4283

    work page 2021

  8. [8]

    Characterizing the temporal dynamics of universal speech representations for generalizable deepfake detection

    Y . Zhu, S. Powar, and T. H. Falk, “Characterizing the temporal dynamics of universal speech representations for generalizable deepfake detection,” 2023. [Online]. Available: https://arxiv.org/abs/2309.08099

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

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

    X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. H. Kinnunen, N. Evans, K. A. Lee, and J. Yamagishi, “ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 1–8

    work page 2024

  10. [10]

    Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization,”International Journal of Computer Vision, vol. 128, no. 2, p. 336–359, Oct. 2019. [Online]. Available: http://dx.doi.org/10.1007/s11263-019-01228-7

    work page doi:10.1007/s11263-019-01228-7 2019

  11. [11]

    A comparison of features for synthetic speech detection,

    M. Sahidullah, T. Kinnunen, and C. Hanilc ¸i, “A comparison of features for synthetic speech detection,” inInterspeech 2015, 2015, pp. 2087– 2091

    work page 2015

  12. [12]

    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. [Online]. Available: https://doi.org/10.3390/s25071989

    work page doi:10.3390/s25071989 1989

  13. [13]

    Investigating self-supervised front ends for speech spoofing countermeasures,

    X. Wang and J. Yamagishi, “Investigating self-supervised front ends for speech spoofing countermeasures,” inThe Speaker and Language Recognition Workshop (Odyssey 2022). ISCA, 2022

    work page 2022

  14. [14]

    Automatic Speaker Verification Spoofing and Deepfake Detection Us- ing Wav2vec 2.0 and Data Augmentation,

    H. Tak, M. Todisco, X. Wang, J. weon Jung, J. Yamagishi, and N. Evans, “Automatic Speaker Verification Spoofing and Deepfake Detection Us- ing Wav2vec 2.0 and Data Augmentation,” inThe Speaker and Language Recognition Workshop (Odyssey 2022), 2022, pp. 112–119

    work page 2022

  15. [15]

    Real-world audio deepfake detection using ssl-based speech models and diverse training data,

    K. Sch ¨afer and M. Neu, “Real-world audio deepfake detection using ssl-based speech models and diverse training data,” in2025 IEEE 37th International Conference on Tools with Artificial Intelligence (ICTAI), 2025, pp. 1465–1469

    work page 2025

  16. [16]

    Speech df arena: A leaderboard for speech deepfake detection models,

    S. Dowerah, A. Kulkarni, A. Kulkarni, H. M. Tran, J. Kalda, A. Fe- dorchenko, B. Fauve, D. Lolive, T. Alum ¨ae, and M. Magimai.-Doss, “Speech df arena: A leaderboard for speech deepfake detection models,” IEEE Open Journal of Signal Processing, vol. 7, pp. 73–81, 2026

    work page 2026

  17. [17]

    RISE: randomized input sampling for explanation of black-box models,

    V . Petsiuk, A. Das, and K. Saenko, “RISE: randomized input sampling for explanation of black-box models,” inBritish Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018. BMV A Press, 2018, p. 151. [Online]. Available: http://bmvc2018.org/contents/papers/1064.pdf

    work page 2018

  18. [18]

    Explainable artificial intelligence: A comprehensive review,

    D. Minh, H. X. Wang, Y . F. Li, and T. N. Nguyen, “Explainable artificial intelligence: A comprehensive review,”Artificial Intelligence Review, vol. 55, no. 5, pp. 3503–3568, 2022

    work page 2022

  19. [19]

    Why Should I Trust You ?

    M. T. Ribeiro, S. Singh, and C. Guestrin, “”why should i trust you?”: Explaining the predictions of any classifier,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’16. New York, NY , USA: Association for Computing Machinery, 2016, p. 1135–1144. [Online]. Available: https://doi.org/10.1145/29...

    work page doi:10.1145/2939672.2939778 2016

  20. [20]

    Explanations for automatic speech recognition,

    X. lan Wu, P. Bell, and A. Rajan, “Explanations for automatic speech recognition,”ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5,

    work page 2023

  21. [21]

    Available: https://api.semanticscholar.org/CorpusID: 257232550

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 257232550

    work page

  22. [22]

    Explainable Attribute-Based Speaker Verification

    X. Wu, C. Luu, P. Bell, and A. Rajan, “Explainable attribute-based speaker verification,” 2024. [Online]. Available: https://arxiv.org/abs/ 2405.19796

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    A systematic review of interpretability and explainability for speech emotion features in automatic speech emotion recognition,

    H. M. Jayasinghe, K. W. Wong, and A. Nugaliyadde, “A systematic review of interpretability and explainability for speech emotion features in automatic speech emotion recognition,”Pattern Recognition, vol. 171, p. 112122, 2026

    work page 2026

  24. [24]

    Explaining deep learning models for spoofing and deepfake detection with shapley additive explanations,

    W. Ge, J. Patino, M. Todisco, and N. Evans, “Explaining deep learning models for spoofing and deepfake detection with shapley additive explanations,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6387– 6391

    work page 2022

  25. [25]

    Sanity checks for saliency maps,

    J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, “Sanity checks for saliency maps,”Advances in neural information processing systems, vol. 31, 2018

    work page 2018

  26. [26]

    Attention is not Explanation,

    S. Jain and B. C. Wallace, “Attention is not Explanation,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), J. Burstein, C. Doran, and T. Solorio, Eds. Minneapolis, Minnesota: Association for Computational Linguistics, Jun. 2...

    work page 2019

  27. [27]

    How to Manipulate CNNs to Make Them Lie: the GradCAM Case

    T. Viering, Z. Wang, M. Loog, and E. Eisemann, “How to Manipulate CNNs to Make Them Lie: the GradCAM Case,” 2019. [Online]. Available: https://arxiv.org/abs/1907.10901

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  28. [28]

    Phonetically explainable speech deepfake detection,

    M. Chhibber, J. Mishra, and T. H. Kinnunen, “Phonetically explainable speech deepfake detection,” 2026. [Online]. Available: https://arxiv.org/abs/2606.15454

    work page arXiv 2026

  29. [29]

    Comprehension of synthetic speech produced by rule: A review and theoretical interpretation,

    S. A. Duffy and D. B. Pisoni, “Comprehension of synthetic speech produced by rule: A review and theoretical interpretation,”Language and Speech, vol. 35, no. 4, pp. 351–389, 1992

    work page 1992

  30. [30]

    Review of text-to-speech conversion for english,

    D. H. Klatt, “Review of text-to-speech conversion for english,”The Journal of the Acoustical Society of America, vol. 82, no. 3, pp. 737– 793, 1987

    work page 1987

  31. [31]

    Forensic deepfake audio detection using segmental speech features,

    T. Yang, C. Sun, S. Lyu, and P. Rose, “Forensic deepfake audio detection using segmental speech features,”Forensic Science International, vol. 379, p. 112768, 2026

    work page 2026

  32. [32]

    Comparison of the ITU-t p.85 standard to other methods for the evaluation of text-to-speech systems,

    D. Sityaev, K. Knill, and T. Burrows, “Comparison of the ITU-t p.85 standard to other methods for the evaluation of text-to-speech systems,” inInterspeech 2006, 2006, pp. paper 1233–Tue2WeO.3

    work page 2006

  33. [33]

    Concatenated phoneme models for text-variable speaker recognition,

    “Concatenated phoneme models for text-variable speaker recognition,” in1993 IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 2. IEEE, 1993, pp. 391–394

    work page 1993

  34. [34]

    Speaker-individuality in suprasegmental temporal features: Implications for forensic voice comparison,

    A. Leemann, M.-J. Kolly, and V . Dellwo, “Speaker-individuality in suprasegmental temporal features: Implications for forensic voice comparison,”Forensic Science International, vol. 238, pp. 59–67, 2014. [Online]. Available: https://doi.org/10.1016/j.forsciint.2014.02.019

    work page doi:10.1016/j.forsciint.2014.02.019 2014

  35. [35]

    Phoneme specific modelling and scoring techniques for anti spoofing system,

    G. Suthokumar, K. Sriskandaraja, V . Sethu, C. Wijenayake, and E. Am- bikairajah, “Phoneme specific modelling and scoring techniques for anti spoofing system,” inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp. 6106–6110

    work page 2019

  36. [36]

    Phoneme Discretized Saliency Maps for Explainable Detection of AI-Generated V oice,

    S. Gupta, M. Ravanelli, P. Germain, and C. Subakan, “Phoneme Discretized Saliency Maps for Explainable Detection of AI-Generated V oice,” inInterspeech 2024, 2024, pp. 3295–3299

    work page 2024

  37. [37]

    Robust speech recognition via large-scale weak supervision,

    A. Radfordet al., “Robust speech recognition via large-scale weak supervision,” inProceedings of the 40th International Conference on Machine Learning, ser. ICML’23. JMLR.org, 2023

    work page 2023

  38. [38]

    BFA: Real- time Multilingual Text-to-speech Forced Alignment,

    A. Rehman, J. Cai, J.-J. Zhang, and X. Yang, “BFA: Real- time Multilingual Text-to-speech Forced Alignment,” 2025. [Online]. Available: https://arxiv.org/abs/2509.23147

    work page arXiv 2025

  39. [39]

    Cupe: Contextless universal phoneme encoder for language-agnostic speech processing,

    A. Rehman, J.-J. Zhang, and X. Yang, “Cupe: Contextless universal phoneme encoder for language-agnostic speech processing,” inPro- ceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP 2025), ICNLSP. International Conference on Natural Language and Speech Processing, 2025

    work page 2025

  40. [40]

    Journal of the American Statistical Association47(260), 583–621 (1952)

    W. H. Kruskal and W. A. Wallis, “Use of ranks in one- criterion variance analysis,”Journal of the American Statistical Association, vol. 47, no. 260, pp. 583–621, 1952. [Online]. Available: https://doi.org/10.1080/01621459.1952.10483441

    work page doi:10.1080/01621459.1952.10483441 1952

  41. [41]

    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

    work page 1995

  42. [42]

    Effect sizes for nonparametric tests,

    F. Fiel Peres, “Effect sizes for nonparametric tests,”Biochemia Medica, vol. 36, no. 1, p. 010101, 2026. [Online]. Available: https://doi.org/10.11613/BM.2026.010101

    work page doi:10.11613/bm.2026.010101 2026

  43. [43]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProceedings of the 3rd International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015, poster. [Online]. Available: https://arxiv.org/abs/1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  44. [44]

    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,” inInternational conference on machine learning. PMLR, 2022, pp. 2709–2720

    work page 2022

  45. [45]

    Relative occurrence of phonemes in american english,

    J. V . Tobias, “Relative occurrence of phonemes in american english,” The Journal of the Acoustical Society of America, vol. 31, no. 5, pp. 631–631, 1959

    work page 1959

  46. [46]

    Journal of the Royal Statistical Society: Series B (Methodological)57(1), 289–300 (1995)

    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, 01 1995. [Online]. Available: https://doi.org/10.1111/j.2517-6161.1995.tb02031.x

    work page doi:10.1111/j.2517-6161.1995.tb02031.x 1995

  47. [47]

    Speech is silver, silence is golden: What do asvspoof- trained models really learn?

    N. M ¨uller, F. Dieckmann, P. Czempin, R. Canals, K. B ¨ottinger, and J. Williams, “Speech is silver, silence is golden: What do asvspoof- trained models really learn?”2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 2021

    work page 2021

  48. [48]

    The impact of silence on speech anti-spoofing,

    Y . Zhang, Z. Li, J. Lu, H. Hua, W. Wang, and P. Zhang, “The impact of silence on speech anti-spoofing,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 3374–3389, 2023

    work page 2023

  49. [49]

    A deeper look at gaussian mixture model based anti-spoofing systems,

    B. Chettri and B. L. Sturm, “A deeper look at gaussian mixture model based anti-spoofing systems,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 5159– 5163

    work page 2018

  50. [50]

    Malafide: a novel adversarial convolutive noise attack against deepfake and spoofing detection systems,

    M. Panariello, W. Ge, H. Tak, M. Todisco, and N. Evans, “Malafide: a novel adversarial convolutive noise attack against deepfake and spoofing detection systems,” inInterspeech 2023, 2023, pp. 2868–2872

    work page 2023

  51. [51]

    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 Asia-Pacific Chapter of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing (AACL-IJCNLP), 2022, pp. 741–751

    work page 2022

  52. [52]

    BigV- GAN: A universal neural vocoder with large-scale training,

    S. gil Lee, W. Ping, B. Ginsburg, B. Catanzaro, and S. Yoon, “BigV- GAN: A universal neural vocoder with large-scale training,” inInterna- tional Conference on Learning Representations (ICLR), 2023

    work page 2023

  53. [53]

    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), N. Calzolari, K. Choukri, C. Cieri, T. Declerck, S. Goggi, K. Hasida, H. Isahara, B. Maegaard, J. Mariani, H. Mazo, A. ...

    work page 2018