Pith. sign in

REVIEW 2 major objections 6 minor 34 references

A new score reveals when audio deepfake detectors fail because of speaker identity, not synthesis artifacts.

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 · deepseek-v4-flash

2026-08-01 06:34 UTC pith:M2XEKTLS

load-bearing objection A useful new per-utterance diagnostic for speaker sensitivity in deepfake detectors, with a solid error-prediction core, but the identity-specific interpretation is not yet cleanly separated from boundary proximity. the 2 major comments →

arxiv 2607.21820 v1 pith:M2XEKTLS submitted 2026-07-23 cs.SD cs.AIcs.CRcs.LG

Probing Speaker Identity Sensitivity in Audio Deepfake Detectors

Daniyal Kabir Dar , Arun Ross This is my paper
classification cs.SD cs.AIcs.CRcs.LG
keywords audio deepfake detectionspeaker identityidentity sensitivity scoregeneralization gapvoice conversionASVspooflabel-free diagnosticuncertainty quantification
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 tries to establish that audio deepfake detectors partly succeed by leaning on who is speaking, not just on how the audio was synthesized, and that this habit can be measured per utterance. It introduces the Identity Sensitivity Score (ISS), a label-free number computed by re-scoring the same audio under many different speaker contexts and taking the spread of those scores. The paper shows wrong predictions have ISS 29–52 times higher than right ones, that ISS predicts misclassification with AUC up to 0.954, and that voice-converted versions of high-ISS utterances shift detector scores 19–30 times more than low-ISS ones. These results support ISS as an inference-time diagnostic for speaker-dependent failures, and they point at identity shortcuts as a contributor to the large cross-dataset error jumps seen in this area.

Core claim

The central claim is that a deepfake detector's score on a single utterance is partly a function of the speaker's identity, and that this identity-dependence is measurable without ground-truth labels. The paper's Identity Sensitivity Score (ISS) does this by computing p_fake(A|S_i)=σ(ℓ(A)+α·cos(A,S_i)) for K enrolled speakers and taking the interquartile range of those scores. On ASVspoof 2019/2021 LA with AASIST and RawNet2, incorrect utterances show 29–52× higher ISS (AUC up to 0.954), and voice conversion moves high-ISS utterances' scores 19.2–30.7× more than low-ISS ones. The paper interprets this as evidence that identity sensitivity is real detector behavior, not just a proxy for confi

What carries the argument

The Identity Sensitivity Score (ISS): for a query utterance A, take the detector's log-odds ℓ(A), add α times the cosine similarity between A's speaker embedding (from a pretrained ECAPA-TDNN) and each of K enrolled speakers' mean embeddings, pass each sum through the sigmoid, and record the interquartile range of the resulting K probabilities. The logit-space addition is a Bayesian score-fusion-style probe that leaves the audio untouched; IQR makes the spread robust to outlier speakers. ISS is doing the work of converting a single scalar score into a family of identity-conditioned scores, so that identity-driven instability becomes visible as spread.

Load-bearing premise

The load-bearing premise is that comparing high-ISS to low-ISS utterances isolates identity sensitivity, but those groups are not matched for distance to the decision boundary, and the sigmoid slope is steepest near p=0.5, so the large voice-conversion response could partly reflect boundary proximity rather than identity.

What would settle it

Rematch the voice-conversion groups so high- and low-ISS utterances have identical distributions of |p_fake−0.5| (or identical entropy, margin, or confidence); if the 19–30× differential collapses or drops sharply, ISS is largely a confidence proxy rather than an identity measure.

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

If this is right

  • At inference, utterances with ISS above a threshold can be flagged for human review without any labels, with AUC up to 0.954 for predicting misclassification.
  • ISS monitors distribution shift: AASIST's 21× EER jump from 2019 to 2021 LA coincides with a 24× amplification of the ISS separation ratio, so rising ISS can signal that a deployment environment is drifting.
  • Attack-type analysis shows TTS and VC errors carry strong ISS elevation, while RawNet2's hybrid-attack errors show none (ratio 0.2×), letting practitioners distinguish identity-driven failures from acoustic-artifact failures.
  • Because ISS is complementary to entropy, margin, and confidence, it can be combined with those baselines rather than replace them for a fuller picture of why a prediction is uncertain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If identity sensitivity is a learned shortcut, training on data where speaker identity is partially unlinked from the genuine/synthetic label (for example, pairing each synthesis system with many speakers and balancing speaker demographics) should reduce both ISS and cross-dataset EER; this is a testable intervention the paper notes but does not run.
  • ISS depends on the external embedding space of ECAPA-TDNN; applying the same probe with a different speaker encoder, or with the detector's own internal representations, would reveal whether the measured identity sensitivity is a property of the detector or of the probe's notion of identity.
  • The logit-space identity perturbation is an assessment, not a cause: nothing in the paper shows that removing identity sensitivity would improve generalization, only that the two are correlated. A training-level ablation that measures ISS before and after identity-invariant regularization would close that gap.
  • Because ISS requires only one extra forward pass through a speaker encoder, it could be repurposed as a regularizer during training, penalizing high-ISS predictions to discourage identity shortcuts rather than merely diagnosing them.

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 / 6 minor

Summary. The paper proposes Identity Sensitivity Score (ISS), a per-utterance diagnostic intended to quantify how much an audio deepfake detector's output depends on speaker identity. ISS is computed by taking the logit of the detector's p_fake, adding α·cos(A,S_i) for K alternative speaker prototypes from an ECAPA-TDNN embedding space, and measuring the interquartile range of the resulting sigmoid probabilities. The authors report that incorrectly classified utterances have 29–52× higher ISS than correctly classified ones, that ISS predicts misclassification with AUC up to 0.954, and that voice-converted versions of high-ISS utterances shift the detector score 19–30× more than low-ISS utterances. They interpret the voice-conversion result as evidence that ISS captures identity sensitivity rather than being a proxy for prediction confidence.

Significance. If the interpretation holds, ISS would be a useful label-free, inference-time diagnostic for speaker-dependent failure analysis in deepfake detection. The paper has genuine strengths: a clear problem formulation, a conceptually interesting logit-space perturbation, two architecturally distinct detectors, two datasets, an honest comparison to entropy/margin/confidence baselines, ablation over K, and a breakdown by attack type. The RawNet2 hybrid-attack result (ISS ratio 0.2×, p=1.00) is a nice example of selectivity. However, the central validation — the voice-conversion experiment — is confounded by decision-boundary proximity, so the paper's main interpretive claim is not currently established. The issue is fixable with matched-group or covariate-controlled analysis, which is why I recommend major revision rather than rejection.

major comments (2)
  1. [§4.4, Table 4; also §3.5] The voice-conversion validation does not control for the base score's distance to the decision boundary. For small α, ISS(A) ≈ α·IQR_i[cos(A,S_i)]·σ(ℓ)(1−σ(ℓ)), so high-ISS utterances are preferentially those with p_fake near 0.5. The measured Δscore = |p_fake(A_conv)−p_fake(A_src)| is similarly bounded by the sigmoid slope. Thus the 19.2×/30.7× differential between high- and low-ISS groups may be a slope artifact rather than evidence of identity-specific sensitivity. The acknowledgment in §3.5 that ISS is coupled to boundary proximity does not remove this confound. Please re-run the VC experiment with high- and low-ISS groups matched on p_fake (e.g., stratified within narrow p_fake bins), or include base p_fake as a covariate, and report the p_fake distributions of both groups. The attack-type results in §5.6 are suggestive but do not resolve this because those groups are defined by err
  2. [§3.4] ISS as defined requires sampling K alternative speakers while 'excluding A's true speaker.' At inference time, the true speaker identity is generally not known; if the query speaker is one of the enrolled speakers, this exclusion requires exactly the kind of label the method claims to avoid. The paper should clarify whether the evaluation speakers are disjoint from the enrollment pool. If they are disjoint, the exclusion step is vacuous and the method is only validated for out-of-enrollment speakers; if they are not disjoint, the label-free claim is problematic. Either way, the deployment scenario needs to be stated precisely.
minor comments (6)
  1. [§4.3, Table 7] The reported ISS ratios and AUCs use α=5.0, which is selected on the development set by maximizing the correct/incorrect ISS ratio. This is disclosed, but the paper should also present the evaluation-set results across the α grid (or a nested selection) to show that the headline numbers are not sensitive to the exact selection criterion. The 'bonafide/spoof ordering inverts' criterion is somewhat fragile and worth explaining.
  2. [§2.3] The baseline called 'confidence' is defined as 1−max(p_fake,1−p_fake), which is actually minimized at the boundary, not maximized. Although it is correctly grouped with entropy and margin as a monotone transform of |p_fake−0.5|, the name is confusing; consider calling it 'uncertainty' or defining it as max(p,1−p).
  3. [§4.4] The protocol says '500 utterances per detector' but the description is 250 high-ISS + 250 low-ISS. Please make the counting explicit.
  4. [Table 3] The 'Advantage' row reports differences in AUC percentage points (e.g., +5.3). Please state the units to avoid confusion with ratios.
  5. [§6.1] The p_fake=0.51 example is repeated nearly verbatim from §5.4. One occurrence would be sufficient.
  6. [Table 7 footnote] The dev-set ratios (6,200×–17,700×) versus evaluation ratios (29×) are attributed to class composition. The explanation is plausible but would be more convincing with a brief table of dev/eval ISS_cor and ISS_inc medians.

Circularity Check

0 steps flagged

No circular derivation: ISS is defined from external speaker embeddings and the base detector score, the main empirical claims are held-out, and the VC validation confound is a validity issue rather than a circular reduction.

full rationale

The derivation of ISS (Eqs. 3-4) is self-contained: it combines the detector's base logit with cosine similarities from an external ECAPA-TDNN speaker encoder and defines ISS as an IQR over K alternative-speaker contexts. It does not use error labels in its definition, so the observed 29-52x separation between correct and incorrect utterances and the AUC of 0.954 are empirical findings on held-out evaluation partitions, not algebraic consequences of the definition. The alpha=5 perturbation strength is tuned on the 2019 LA development set by maximizing the correct/incorrect ISS ratio; this is standard hyperparameter selection, and the headline numbers are reported on evaluation partitions, so this is not a fitted parameter being renamed as a prediction. The voice-conversion validation (Sec. 4.4, 5.5) has a genuine confound: high-ISS and low-ISS groups are selected by ISS rank without matching on p_fake, and since ISS = IQR{sigma(ell+alpha cos)} is amplified by the sigmoid slope sigma'(ell)=sigma(ell)(1-sigma(ell)) near p=0.5, the 19-30x Delta-score differential could partly reflect boundary proximity rather than identity. However, this is a validity/confound problem with an experimental test, not a circular derivation: Delta-score = |p_fake(A_conv)-p_fake(A_src)| is an external measurement and is not equal to Eq. (3) by construction. The only self-citation ([5]) is contextual background and is not load-bearing; no uniqueness theorem or prior ansatz is invoked to force the core claim. The authors themselves note in Sec. 6.2 that the VC validation does not establish a causal training-level link. Accordingly, no step in the paper's derivation chain reduces to its own inputs, and the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The metric is built on six assumptions: external embedding as identity proxy, logit-space injection as a valid probe, IQR as sensitivity statistic, the dataset label-identity correlation, transferability of α, and feasibility of true-speaker exclusion at inference. No new physical entities are introduced. The most consequential assumption is that the voice-conversion experiment disambiguates identity from confidence, which is not established because the groups differ in boundary proximity.

free parameters (4)
  • α (logit perturbation strength) = 5.0
    Tuned on the ASVspoof 2019 LA development set by maximizing the correct/incorrect ISS ratio before the bonafide/spoof median ordering inverts (Table 7); all headline ISS ratios and AUCs depend on this value.
  • K (number of alternative speaker contexts) = 10
    Chosen from ablation over K in {3,5,10,20} (Table 6) showing stable ratios; affects the IQR and therefore every ISS value.
  • N (enrollment utterances per speaker prototype) = 5
    Used to build ECAPA-TDNN prototypes (Eq. 1); no sensitivity analysis for N is provided.
  • VC high/low ISS group thresholds = AASIST ISS>0.05 / <0.01; RawNet2 medians 3.1e-3 vs 5.8e-6
    Rank-based thresholds define the high- and low-ISS groups in the voice-conversion validation (Section 4.4); groups are not matched on base score, confounding identity with boundary proximity.
axioms (6)
  • domain assumption ECAPA-TDNN cosine similarity is a valid proxy for speaker identity as relevant to the deepfake detector's decision.
    Used in Eq. 3 to define identity-conditioned scores; the paper itself notes in Section 6.2 that ECAPA-TDNN's identity space may not align with the detector's internal representations.
  • ad hoc to paper Adding α·cos(A,S_i) in logit space produces meaningful identity-conditioned probabilities.
    Section 3.3: the audio is never modified and the detector's internals are not used, so the probe's validity rests on the Bayesian-fusion analogy [2] rather than on direct evidence about the detector.
  • ad hoc to paper IQR of the identity-conditioned scores is the right measure of identity sensitivity.
    Section 3.4: robustness to outliers motivates IQR over variance, but no formal argument connects IQR to identity reliance in the detector.
  • domain assumption ASVspoof 2019 LA training data correlates speaker identity with genuine/synthetic labels.
    Section 2.1 argues synthetic samples are produced by systems trained on the real speakers in the dataset; the entire motivation depends on this correlation.
  • domain assumption α tuned on a labeled development set transfers to the unlabeled deployment setting.
    Section 4.3 acknowledges α tuning requires labels; transferability across domains/datasets is assumed rather than demonstrated.
  • ad hoc to paper The true speaker of a query can be excluded from the K sampled contexts at inference.
    Section 3.4 says 'excluding A's true speaker,' but label-free inference does not provide the true speaker identity; this contradicts the label-free claim.

pith-pipeline@v1.3.0-alltime-deepseek · 12192 in / 16386 out tokens · 158936 ms · 2026-08-01T06:34:47.722330+00:00 · methodology

0 comments
read the original abstract

Audio deepfake detectors are trained to distinguish genuine speech from synthetic speech and often perform well on standard benchmarks. Yet the same detector that achieves less than 1% error on one dataset can see its error rate increase twentyfold when evaluated on a different dataset. We argue that one contributing factor is speaker-identity reliance: standard training corpora correlate speaker identity with the genuine/synthetic label, allowing detectors to partially rely on speaker-related cues rather than synthesis artifacts alone. We propose the Identity Sensitivity Score (ISS), a per-utterance diagnostic that quantifies how much a detector's output changes across different speaker identity contexts. ISS requires no ground-truth labels at inference time and can be computed from the detector score and a pool of reference speaker examples. Across two detectors and two datasets, incorrectly classified utterances have ISS scores 29 to 52 times higher than correctly classified utterances, and ISS alone predicts misclassification with area-under-curve (AUC) up to 0.954. To test whether ISS actually captures identity-sensitive behavior rather than serving only as a proxy for prediction confidence, we apply voice conversion to 500 utterances and measure the resulting detector-score shift. Utterances flagged as identity-sensitive by ISS respond 19 to 30 times more strongly to this manipulation than utterances flagged as stable. These results position ISS as a practical inference-time diagnostic for speaker-dependent failure analysis in audio deepfake detection.

Figures

Figures reproduced from arXiv: 2607.21820 by Arun Ross, Daniyal Kabir Dar.

Figure 1
Figure 1. Figure 1: The ISS concept. Top: A standard detector produces a single score pfake with no indication of whether the prediction is reliable — the same score could reflect synthesis artifacts or identity-dependent behavior. Bottom: ISS evaluates the same audio under K varied speaker identity contexts (S1, S2, S3) via ECAPA-TDNN prototypes and measures the score spread via IQR (interquartile range). A high spread (ISS … view at source ↗
Figure 2
Figure 2. Figure 2: ISS distributions for correctly and incorrectly classified [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ROC curves for ISS as a label-free predictor of misclas [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

34 extracted references · 5 linked inside Pith

  1. [1]

    ASVspoof 2021 baseline systems

    ASVspoof Challenge. ASVspoof 2021 baseline systems. https : / / github . com / asvspoof - challenge / 2021, 2021

    2021

  2. [2]

    Br ¨ummer and J

    N. Br ¨ummer and J. du Preez. Application-independent eval- uation of speaker detection.Computer Speech & Language, 20(2–3):230–275, 2006

    2006

  3. [3]

    Y . Chen, J. Yi, J. Xue, C. Wang, X. Zhang, S. Dong, S. Zeng, J. Tao, L. Zhao, and C. Fan. RawBMamba: End-to-end bidirectional state space model for audio deepfake detection. arXiv preprint arXiv:2406.06086, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Cheng, H

    Z. Cheng, H. Yang, Y . Xiong, and X. Hu. Explainable AI for forensic speech authentication within cognitive and compu- tational neuroscience.Frontiers in Neuroscience, 19, 2025

    2025

  5. [5]

    D. K. Dar, Q. Yan, L. Xiao, and A. Ross. Impact of phonetics on speaker identity in adversarial voice attack. InProceed- ings IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 13462–13466, 2026

    2026

  6. [6]

    Delgado, N

    H. Delgado, N. Evans, T. Kinnunen, K. A. Lee, V . Vestman, A. Nautsch, M. Todisco, X. Wang, M. Sahidullah, J. Yam- agishi, and D. A. Reynolds. ASVspoof 2021: Automatic speaker verification spoofing and countermeasures challenge evaluation plan. InProceedings ASVspoof Workshop, 2021

    2021

  7. [7]

    Desplanques, J

    B. Desplanques, J. Thienpondt, and K. Demuynck. ECAPA- TDNN: Emphasized channel attention, propagation and ag- gregation in TDNN based speaker verification. InProceed- ings Interspeech, pages 3830–3834, 2020

    2020

  8. [8]

    Frank and L

    J. Frank and L. Sch ¨onherr. WaveFake: A data set to facilitate audio deepfake detection. InProceedings NeurIPS Datasets and Benchmarks Track, 2021

    2021

  9. [9]

    Gal and Z

    Y . Gal and Z. Ghahramani. Dropout as a Bayesian approxi- mation: Representing model uncertainty in deep learning. In Proceedings International Conference on Machine Learning (ICML), pages 1050–1059, 2016

    2016

  10. [10]

    H. Gu, J. Yi, C. Wang, Y . Ren, J. Tao, X. Yan, Y . Chen, and X. Zhang. Utilizing speaker profiles for impersonation audio detection. InProceedings 32nd ACM International Conference on Multimedia, pages 1961–1970, 2024

    1961

  11. [11]

    P. Kawa, M. Plata, and P. Syga. Improved deepfake detection using whisper features. InProceedings Interspeech, pages 4009–4013, 2023

    2023

  12. [12]

    Khanjani, G

    Z. Khanjani, G. Watson, and V . P. Janeja. Audio deepfakes: A survey.Frontiers in Big Data, 5:1001063, 2023

    2023

  13. [13]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep en- sembles.Advances in Neural Information Processing Sys- tems, 2017

    2017

  14. [14]

    J. Li, W. Lin, L. Chen, and Z. Wu. FreeVC: Towards high- quality text-free one-shot voice conversion. InProceedings IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023

    2023

  15. [15]

    X. Liu, M. Liu, L. Wang, K. A. Lee, H. Zhang, and J. Dang. Leveraging positional-related local-global dependency for synthetic speech detection. InProceedings IEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 1–5, 2023

    2023

  16. [16]

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Del- gado, T. Kinnunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautsch, and K. A. Lee. ASVspoof 2021: To- wards spoofed and deepfake speech detection in the wild. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2507–2522, 2023

    2021

  17. [17]

    M ¨uller, P

    N. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B¨ottinger. Does audio deepfake detection generalize? In Proceedings Interspeech, pages 2783–2787, 2022

    2022

  18. [18]

    Nagrani, J

    A. Nagrani, J. S. Chung, and A. Zisserman. V oxCeleb: A large-scale speaker identification dataset. InProceedings In- terspeech, pages 2616–2620, 2017

    2017

  19. [19]

    J. Pan, S. Nie, H. Zhang, S. He, K. Zhang, S. Liang, X. Zhang, and J. Tao. Speaker recognition-assisted robust audio deepfake detection. InProceedings Interspeech, pages 4202–4206, 2022

    2022

  20. [20]

    Pianese, D

    A. Pianese, D. Cozzolino, G. Poggi, and L. Verdoliva. Deep- fake audio detection by speaker verification. InProceedings IEEE International Workshop on Information Forensics and Security (WIFS), pages 1–6, 2022

    2022

  21. [21]

    Ravanelli, T

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cor- nell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, et al. Speechbrain: A general-purpose speech toolkit, 2021

    2021

  22. [22]

    Sisman, J

    B. Sisman, J. Yamagishi, S. King, and H. Li. An overview of voice conversion and its challenges: From statistical mod- eling to deep learning.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:132–157, 2021

    2021

  23. [23]

    Tak, J.-w

    H. Tak, J.-w. Jung, J. Patino, M. Kamble, M. Todisco, and N. Evans. End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detection.arXiv preprint arXiv:2107.12710, 2021

    Pith/arXiv arXiv 2021

  24. [24]

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

    2021

  25. [25]

    H. Tak, M. Todisco, X. Wang, J. Yamagishi, and N. Evans. Automatic speaker verification spoofing and deepfake de- tection using wav2vec 2.0 and data augmentation. InPro- ceedings The Speaker and Language Recognition Workshop (Odyssey), pages 112–119, 2022

    2022

  26. [26]

    X. Tan, T. Qin, F. Soong, and T.-Y . Liu. A survey on neural speech synthesis.arXiv preprint arXiv:2106.15561, 2021

    Pith/arXiv arXiv 2021

  27. [27]

    Todisco, X

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Del- gado, A. Nautsch, J. Yamagishi, N. Evans, T. Kinnunen, and K. A. Lee. ASVspoof 2019: Future horizons in spoofed and fake audio detection. InProceedings Interspeech, pages 1008–1012, 2019

    2019

  28. [28]

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . Vestman, T. Kinnunen, K. A. Lee, L. Juvela, P. Alku, Y .-H. Peng, H.-T. Hwang, Y . Tsao, H.-M. Wang, S. L. Maguer, M. Becker, F. Henderson, R. Clark, Y . Zhang, and Q. Wang. ASVspoof 2019: A large- scale public database of synthesized, converted and replayed speech.Com...

    2019

  29. [29]

    weon Jung, S

    J. weon Jung, S. bin Kim, H. jin Shim, J. ho Kim, and H.-J. Yu. RawNet: Advanced end-to-end deep neural network us- ing raw waveforms for text-independent speaker verification. InProceedings Interspeech, pages 1268–1272, 2019

    2019

  30. [30]

    weon Jung, H.-S

    J. weon Jung, H.-S. Heo, H. Tak, H. jin Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans. AASIST: Audio anti- spoofing using integrated spectro-temporal graph attention networks. InProceedings IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6367–6371, 2022

    2022

  31. [31]

    J. Yi, Y . Bai, J. Tao, H. Ma, Z. Tian, C. Wang, T. Wang, and R. Fu. ADD 2022: The first audio deep fake detection challenge. InProceedings IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 9216–9220, 2022

    2022

  32. [32]

    J. Yi, J. Tao, R. Fu, X. Yan, C. Wang, T. Wang, C. Y . Zhang, X. Zhang, Y . Zhao, Y . Ren, L. Xu, J. Zhou, H. Gu, Z. Wen, S. Liang, Z. Lian, S. Nie, and H. Li. ADD 2023: The second audio deepfake detection challenge.arXiv preprint arXiv:2305.13774, 2023

    Pith/arXiv arXiv 2023

  33. [33]

    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

  34. [34]

    Zhang, S

    Q. Zhang, S. Wen, and T. Hu. Audio deepfake detection with self-supervised XLS-R and SLS classifier. InProceedings 32nd ACM International Conference on Multimedia, pages 6765–6773, 2024

    2024