Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Trusted Fake Audio Detection Based on Dirichlet Distribution

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that modeling decision uncertainty with a Dirichlet distribution makes fake audio detectors more accurate and better calibrated.

desk verdict A competent application of evidential deep learning to fake audio detection with consistent but small EER gains, whose headline calibration/trustworthiness claim is undermined by an unfair baseline normalization in Section 4.7. read the letter →

arxiv 2506.02401 v1 pith:LSYHP7MS submitted 2025-06-03 cs.SD cs.MMeess.AS

Chi Ding , Junxiao Xue , Cong Wang , Hao Zhou This is my paper
classification cs.SDcs.MMeess.AS
keywords fakeaudiodetectionuncertaintymodelingDirichletdistributionanti-spoofingevidentialdeeplearningcalibrationASVspooftrustworthiness
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that fake audio detectors become both more accurate and more trustworthy when each decision is accompanied by a Dirichlet-based uncertainty estimate. The proposed method takes an existing anti-spoofing network, swaps its softmax for a non-negative evidence layer, and trains it with an evidential loss so that the network outputs Dirichlet parameters rather than plain scores. From these parameters the method derives class beliefs, predicted probabilities, and an overall uncertainty $u$ for every input. Across the ASVspoof 2019 LA, 2021 LA, and 2021 DF evaluation sets, the trusted versions of AASIST, RawNet2, and RawGAT-ST attain lower error rates (EER and min t-DCF) than their baselines, while also improving calibration as measured by adaptive expected calibration error (aECE) and prediction confidence consistency (PCC). The practical payoff is a detector that can flag uncertain samples for human review instead of committing to a possibly wrong label.

What carries the argument

The Dirichlet distribution is a probability distribution over probability vectors, used here as the conjugate prior for the binary classification distribution. The evidential network obtains evidence by replacing the final softmax with a softplus (non-negative) activation, giving $\alpha_i = e_i + 1$. The strength $S = \sum_i(e_i+1)$ controls uncertainty through $u = K/S$ (with $K=2$ classes), while the predicted probability is the Dirichlet mean $p_i = \alpha_i/S$. The loss combines the integrated cross-entropy (digamma form) with a KL divergence term whose weight $\lambda_t$ grows during training, so the network learns to concentrate evidence on the correct class. This machinery is what carries the argument: it turns any classifier into an evidential classifier with a built-in, trainable uncertainty estimate.

What would settle it

Recalibrate the original AASIST, RawNet2, and RawGAT-ST scores on a held-out split using temperature scaling or isotonic regression, then recompute aECE and PCC against the trusted models; if the recalibrated baselines match or beat the trusted models, the paper's calibration claim collapses. As a second check, test on a held-out attack type and see whether the trusted models' uncertainty $u$ is actually higher on the samples they get wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that modeling the second-order probability of a spoof/bonafide decision with a Dirichlet distribution converts a standard fake audio detector into a 'trusted' detector that is both more accurate and better calibrated. For each input, the evidential network produces non-negative evidence $e_1, e_2$; these set the Dirichlet parameters $\alpha_i = e_i + 1$, from which the belief masses $b_i = e_i/S$, the class probabilities $p_i = \alpha_i/\sum_j \alpha_j$, and the uncertainty $u = 2/S$ are computed. Training uses a Bayes-risk cross-entropy loss plus an annealed KL divergence that suppresses evidence for the wrong class. On ASVspoof 2019 LA, 2021 LA, and 2021 DF, the trusted AASIST, RawNet2, and RawGAT-ST models report lower EER and min t-DCF than their reproduced baselines, and average aECE drops by 89.3%, 71.6%, and 90.2% respectively. The paper also shows that accuracy declines as $u$ increases, which supports using $u$ as a deferral signal.

Load-bearing premise

The reported calibration gains depend on treating the raw scores of the baseline detectors, after simple normalization to [0,1], as if they were valid confidence probabilities; if that normalization is not a legitimate confidence measure, the large aECE reductions claimed in Section 4.7 are artifacts.

Editorial extensions

If this is right

  • Converting an existing detector to evidential form requires only a new final layer and loss, so the recipe can be applied to any current anti-spoofing backbone.
  • The uncertainty output $u$ gives a concrete deferral rule: send high-$u$ inputs to human review or a more powerful model, which is exactly what a trusted detection system needs.
  • The reported gains hold on the hardest conditions tested, including telephony-distorted 2021 LA and codec-degraded 2021 DF, not just the matched 2019 LA condition.
  • Because the model produces probabilities and uncertainty together in one forward pass, no extra inference-time sampling or ensembling is needed to obtain confidence estimates.

Reading between the lines

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

  • An implication the paper leaves implicit is that only aleatoric (data) uncertainty is modeled; epistemic uncertainty about unseen attack algorithms is not captured, so the trusted models may still be overconfident on genuinely novel spoofing techniques.
  • A stricter benchmark for this line of work would recalibrate baselines with post-hoc methods before comparing aECE, since raw scores are not calibrated probabilities.
  • The same evidence-plus-Dirichlet head could be attached to self-supervised speech encoders or large pretrained audio models, since the modification is confined to the classification head and loss.
  • A natural extension would be to threshold on $u$ and measure the accuracy of accepted samples per spoofing attack, producing a deferral operating curve that the paper does not report.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a 'trusted' fake audio detection method that augments three existing backbones (AASIST, RawNet2, RawGAT-ST) with evidential deep learning based on the Dirichlet distribution. The final softmax layer is replaced with softplus to produce evidence, the Dirichlet parameters are derived from that evidence, and training uses the evidential cross-entropy loss with a KL regularization term. Experiments on ASVspoof 2019 LA, ASVspoof 2021 LA, and ASVspoof 2021 DF report lower EER and min t-DCF for the trusted versions of all three backbones, as well as much lower adaptive Expected Calibration Error (aECE) and a qualitative uncertainty-accuracy relationship. The central claim is that the proposed models improve accuracy, robustness, and trustworthiness simultaneously.

Significance. If the results hold, the paper would demonstrate a simple, model-agnostic way to inject uncertainty quantification into strong anti-spoofing detectors, with both accuracy and calibration improvements on standard benchmarks. The strengths are the breadth of the evaluation (three backbones, three datasets, per-attack breakdowns) and the use of an established evidential learning framework. However, the main trustworthiness claim rests on a calibration comparison whose baseline normalization is not a valid confidence measure, and the reported accuracy gains are small and lack statistical support. As presented, the evidence is not yet sufficient to support the abstract's and conclusion's claims of significant improvements in trustworthiness metrics.

major comments (4)
  1. [Section 4.7, Table 5, Eq. (10)] The baseline aECE comparison is not a valid calibration comparison. The paper states that baseline confidence is obtained by normalizing the original model's output scores to [0,1]; for AASIST and RawGAT-ST those scores are pre-softmax logits, and for RawNet2 they are log-softmax outputs. Min-max normalization is an arbitrary monotone transformation that does not produce a probability or a confidence measure, so feeding those values into Eq. (10) measures something other than calibration error. Consequently, the claimed relative aECE reductions of 89.3%, 71.6%, and 90.2%, which underpin the 'trustworthiness' conclusion, may be artifacts of the normalization. Please recompute baseline calibration using a proper posterior (e.g., sigmoid or softmax over logits) and, ideally, include standard calibration baselines such as temperature scaling or Platt scaling, so that the trusted model's Dirichlet-based expected probabilities are compared on a consistent, meaningful scale.
  2. [Tables 2-4 and Section 4.4-4.6] The reported EER and min t-DCF improvements are small and are presented without error bars or significance tests. For example, trusted AASIST improves from 8.16% to 7.65% on ASVspoof 2021 LA and from 20.28% to 19.91% on 2021 DF; the latter is a relative improvement of about 1.8%, which is likely within run-to-run variability for these models. The conclusion's wording 'significant improvements' needs statistical support, such as multiple training seeds with standard deviations and a paired significance test (e.g., Wilcoxon signed-rank) over evaluation conditions or bootstrap over trials.
  3. [Section 4.7 and Conclusion] The abstract and conclusion claim improvements in PCC, but no PCC values are reported anywhere in the paper. Section 4.7 says the authors 'calculate the PCC', yet Table 5 only lists aECE values, and Figure 3 shows qualitative calibration plots but no numeric PCC. Either report the PCC numbers in Table 5 (or a new table), or remove the PCC claim from the abstract and conclusion.
  4. [Eq. (9) and Section 4.3] The loss in Eq. (9) includes the regularization weight lambda_t, and the text says it 'can be gradually increased as the training progresses', but no annealing schedule or concrete values are given in Section 4.3. Since lambda_t controls the strength of the KL term and is a free hyperparameter, its omission makes the experiments irreproducible. Please specify the schedule (e.g., the formula or the per-epoch values) and the final value used in all experiments.
minor comments (6)
  1. [Tables 2, 3, 4] The header 'Trusted AAIST' is a typo for 'Trusted AASIST'.
  2. [Section 4.7] The phrase 'average aEERs' should read 'average aECE'.
  3. [Figure 1 caption] The caption says the steps are 'as described in Section 2', but the method is described in Section 3; please correct the cross-reference.
  4. [Introduction] The contribution list is introduced as 'twofold' but contains three bullets; renumber or rephrase.
  5. [Section 4.4-4.5] The paper notes that trusted models sometimes degrade on individual attacks (e.g., AASIST A09 from 0% to 0.04%, RawNet2 A13 from 1.19% to 1.26%), but does not discuss whether this tradeoff is acceptable for the claimed robustness; a brief analysis would strengthen the narrative.
  6. [Table 1] External baseline rows are not all evaluated under the same training/evaluation protocol as the reproduced models; adding a note or separating reproduced systems from literature-reported numbers would clarify the comparison.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the Dirichlet-evidence method is taken from external prior work, and the accuracy and calibration claims are evaluated on held-out ASVspoof data; the Section 4.7 baseline-normalization issue is a comparison-validity concern rather than a circular derivation.

full rationale

The derivation chain is not circular in the relevant sense. Evidence is produced by replacing the final softmax or log-softmax layer with softplus (Section 4.3), Dirichlet parameters are set as alpha = e + 1 (Eq. 3), beliefs and uncertainty follow Dempster-Shafer theory (Eqs. 2-4), and the training loss (Eqs. 7-9) is the standard evidential deep-learning objective from Sensoy et al. [18], which is external prior work whose assumptions do not include the ASVspoof results. No parameter is fitted to the evaluation sets and then reported as a prediction; EER and min t-DCF are computed on held-out ASVspoof 2019 LA, 2021 LA, and 2021 DF evaluation data, and the aECE and PCC values are computed from the trusted models' outputs on the same held-out data. The only coauthor citation, Ref. [13] (Xue et al.), appears in the general related-work list and is not load-bearing, so at most it is a minor self-citation. The normalization of baseline scores in Section 4.7 is a genuine methodological weakness, because min-max scaling is not a principled confidence measure and may inflate the reported relative aECE reductions, but that is a benchmarking-validity concern rather than a case where a prediction reduces to a fitted input or a self-citation chain. Therefore no specific circular step is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper contributes an application, not new theory. The main assumptions are the validity of the evidential framework on audio and the fairness of the aECE comparison with normalized baselines. The only paper-specific hyperparameter is the annealing lambda_t, which is not specified.

free parameters (1)
  • lambda_t annealing schedule
    Balancing factor between the evidential cross-entropy loss and KL divergence (Eq. 9). The paper says it is gradually increased but gives no values or schedule, so the results depend on an unspecified hyperparameter.
assumptions (3)
  • domain assumption The evidential deep learning loss (Eqs. 7-9) from Sensoy et al. transfers to fake audio detection and yields calibrated uncertainty.
    The paper applies the framework without re-validating its assumptions on audio data.
  • domain assumption The Dirichlet distribution's parameters, derived from network evidence, meaningfully represent belief and uncertainty for binary fake audio detection.
    This is the core postulate of evidential deep learning, inherited from Dempster-Shafer theory.
  • ad hoc to paper The baseline model scores, after min-max normalization to [0,1], serve as valid confidence estimates for computing aECE.
    Section 4.7 uses this comparison to claim trustworthiness improvements; the normalization is not justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Trusted Fake Audio Detection Based on Dirichlet Distribution." pith.science (2026). https://pith.science/paper/LSYHP7MS

@misc{pith2026250602401,
  author       = {Pith},
  title        = {Pith review of: Trusted Fake Audio Detection Based on Dirichlet Distribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LSYHP7MS}},
  note         = {Machine review of arXiv:2506.02401}
}
read the original abstract

With the continuous development of deep learning-based speech conversion and speech synthesis technologies, the cybersecurity problem posed by fake audio has become increasingly serious. Previously proposed models for defending against fake audio have attained remarkable performance. However, they all fall short in modeling the trustworthiness of the decisions made by the models themselves. Based on this, we put forward a plausible fake audio detection approach based on the Dirichlet distribution with the aim of enhancing the reliability of fake audio detection. Specifically, we first generate evidence through a neural network. Uncertainty is then modeled using the Dirichlet distribution. By modeling the belief distribution with the parameters of the Dirichlet distribution, an estimate of uncertainty can be obtained for each decision. Finally, the predicted probabilities and corresponding uncertainty estimates are combined to form the final opinion. On the ASVspoof series dataset (i.e., ASVspoof 2019 LA, ASVspoof 2021 LA, and DF), we conduct a number of comparison experiments to verify the excellent performance of the proposed model in terms of accuracy, robustness, and trustworthiness.

Figures

Figures reproduced from arXiv: 2506.02401 by the authors.

Figure 1
Figure 1. The overall architecture of the proposed trusted fake audio detection method. This method refers to evidence generation, uncertainty modeling based on [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. ASVspoof datasets t-SNE Visualization The ASVspoof 2021 LA evaluation dataset expands on the 2019 LA dataset, containing 181,566 utterances. Unlike its predecessor, this dataset incorporates noise by transmitting each speech sample through various telephone systems, includ￾ing Voice over IP (VoIP) and the Public Switched Telephone Network (PSTN). Consequently, the 2021 LA dataset exhibits a significantly different f… view at source ↗
Figure 3
Figure 3. Calibration for different models on ASVspoof datastes 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: visually illustrates the relationship between the un￾certainty of decisions made by various trusted models and their prediction accuracy. As demonstrated in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

  1. [18]

    Sensoy, L

    M. Sensoy, L. Kaplan, M. Kandemir, Evidential deep learning to quan- tify classification uncertainty, Advances in neural information processing systems 31 (2018)

    work page 2018

  2. [1]

    Ling, S.-Y

    Z.-H. Ling, S.-Y . Kang, H. Zen, A. Senior, M. Schuster, X.-J. Qian, H. M. Meng, L. Deng, Deep learning for acoustic modeling in parametric speech generation: A systematic review of existing techniques and future trends, IEEE Signal Processing Magazine 32 (3) (2015) 35–52

    work page 2015

  3. [2]

    D. Min, D. B. Lee, E. Yang, S. J. Hwang, Meta-stylespeech: Multi- speaker adaptive text-to-speech generation, in: International Conference on Machine Learning, PMLR, 2021, pp. 7748–7759

    work page 2021

  4. [3]

    Theune, E

    M. Theune, E. Klabbers, J.-R. De Pijper, E. Krahmer, J. Odijk, From data to speech: a general approach, Natural Language Engineering 7 (1) (2001) 47–86

    work page 2001

  5. [4]

    J. Kong, J. Kim, J. Bae, Hifi-gan: Generative adversarial networks for ef- ficient and high fidelity speech synthesis, Advances in neural information processing systems 33 (2020) 17022–17033

    work page 2020

  6. [5]

    Gogate, K

    M. Gogate, K. Dashtipour, A. Hussain, Robust real-time audio-visual speech enhancement based on dnn and gan, IEEE Transactions on Ar- tificial Intelligence (2024)

    work page 2024

  7. [6]

    Y . Xiao, K. Shu, H. Zhang, B. Yin, W. S. Cheang, H. Wang, J. Gao, Eggesture: Entropy-guided vector quantized variational autoencoder for co-speech gesture generation, in: Proceedings of the 32nd ACM Interna- tional Conference on Multimedia, 2024, pp. 6113–6122

    work page 2024

  8. [7]

    Evans, C

    Z. Evans, C. Carr, J. Taylor, S. H. Hawley, J. Pons, Fast timing- conditioned latent audio di ffusion, in: Forty-first International Confer- ence on Machine Learning, 2024

    work page 2024

Show all 56 references
  1. [8]

    Chesney, D

    R. Chesney, D. Citron, Deepfakes and the new disinformation war: The coming age of post-truth geopolitics, Foreign Aff. 98 (2019) 147

    2019

  2. [9]

    Borrelli, P

    C. Borrelli, P. Bestagini, F. Antonacci, A. Sarti, S. Tubaro, Synthetic speech detection through short-term and long-term prediction traces, EURASIP Journal on Information Security 2021 (1) (2021) 1–14

    2021

  3. [10]

    Z. Lv, S. Zhang, K. Tang, P. Hu, Fake audio detection based on un- supervised pretraining models, in: ICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2022, pp. 9231–9235

    2022

  4. [11]

    Wu, H.-C

    H. Wu, H.-C. Kuo, N. Zheng, K.-H. Hung, H.-Y . Lee, Y . Tsao, H.-M. Wang, H. Meng, Partially fake audio detection by self-attention-based fake span discovery, in: ICASSP 2022-2022 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2022, ...

    2022

  5. [12]

    H. Ma, J. Yi, J. Tao, Y . Bai, Z. Tian, C. Wang, Continual learning for fake audio detection, arXiv preprint arXiv:2104.07286 (2021). 9

    2021 arXiv

  6. [13]

    J. Xue, H. Zhou, H. Song, B. Wu, L. Shi, Cross-modal information fusion for voice spoofing detection, Speech Communication 147 (2023) 41–50

    2023

  7. [14]

    Gawlikowski, C

    J. Gawlikowski, C. R. N. Tassi, M. Ali, J. Lee, M. Humt, J. Feng, A. Kruspe, R. Triebel, P. Jung, R. Roscher, et al., A survey of uncer- tainty in deep neural networks, Artificial Intelligence Review 56 (Suppl

  8. [15]

    Y . Gal, Z. Ghahramani, Dropout as a bayesian approximation: Represent- ing model uncertainty in deep learning, JMLR.org (2015)

    2015

  9. [16]

    A. G. Wilson, P. Izmailov, Bayesian deep learning and a probabilistic perspective of generalization, Advances in neural information processing systems 33 (2020) 4697–4708

    2020

  10. [17]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable pre- dictive uncertainty estimation using deep ensembles, Advances in neural information processing systems 30 (2017)

    2017

  11. [19]

    X. Wu, R. He, Z. Sun, T. Tan, A light cnn for deep face representation with noisy labels, Journal of IEEE Transactions on Information Forensics and Security 13 (11) (2018) 2884–2896

    2018

  12. [20]

    Cheng, M

    X. Cheng, M. Xu, T. F. Zheng, Replay detection using cqt-based modified group delay feature and resnewt network in asvspoof 2019, in: 2019 Asia- Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), IEEE, 2019, pp. 540–545

    2019

  13. [21]

    Lavrentyeva, S

    G. Lavrentyeva, S. Novoselov, A. Tseren, M. V olkova, A. Gorlanov, A. Kozlov, Stc antispoofing systems for the asvspoof2019 challenge, in: Interspeech, 2019, pp. 1033—-1037

    2019

  14. [22]

    Tomilov, A

    A. Tomilov, A. Svishchev, M. V olkova, A. Chirkovskiy, A. Kondratev, G. Lavrentyeva, Stc antispoofing systems for the asvspoof2021 challenge, in: Proc. ASVspoof 2021 Workshop, 2021, pp. 61–67

    2021

  15. [23]

    T. Chen, E. Khoury, K. Phatak, G. Sivaraman, Pindrop labs’ submission to the asvspoof 2021 challenge, Proc. 2021 edition of the automatic speaker verification and spoofing countermeasures challenge (2021) 89–93

    2021

  16. [24]

    R. Yan, C. Wen, S. Zhou, T. Guo, W. Zou, X. Li, Audio deepfake detec- tion system with neural stitching for add 2022, in: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), IEEE, 2022, pp. 9226–9230

    2022

  17. [25]

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, A. Larcher, End- to-end anti-spoofing with rawnet2, in: ICASSP 2021-2021 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2021, pp. 6369–6373

    2021

  18. [26]

    Ravanelli, Y

    M. Ravanelli, Y . Bengio, Speaker recognition from raw waveform with sincnet, in: 2018 IEEE spoken language technology workshop (SLT), IEEE, 2018, pp. 1021–1028

    2018

  19. [27]

    Tak, J.-w

    H. Tak, J.-w. Jung, J. Patino, M. Todisco, N. Evans, Graph attention net- works for anti-spoofing, arXiv preprint arXiv:2104.03654 (2021)

    2021 arXiv

  20. [28]

    Tak, J.-w

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

    2021 arXiv

  21. [29]

    Jung, H.-S

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.- J. Yu, N. Evans, Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks, in: ICASSP 2022-2022 IEEE interna- tional conference on acoustics, speech and signal processing (ICA...

    2022

  22. [30]

    X. Liu, M. Liu, L. Wang, K. A. Lee, H. Zhang, J. Dang, Leveraging positional-related local-global dependency for synthetic speech detection, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023, pp. 1–5

    2023

  23. [31]

    Z. Lei, H. Yan, C. Liu, Y . Zhou, M. Ma, Gmm-resnet2: Ensemble of group resnet networks for synthetic speech detection, in: ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2024, pp. 12101–12105

    2024

  24. [32]

    C. Liu, X. Xu, F. Xiao, Assd: An ai-synthesized speech detection scheme using whisper feature and types classification, IEEE Transactions on Au- dio, Speech and Language Processing (2025)

    2025

  25. [33]

    C. Guo, G. Pleiss, Y . Sun, K. Q. Weinberger, On calibration of mod- ern neural networks, in: International conference on machine learning, PMLR, 2017, pp. 1321–1330

    2017

  26. [34]

    S. Seo, P. H. Seo, B. Han, Learning for single-shot confidence calibration in deep neural networks through stochastic inferences, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9030–9038

    2019

  27. [35]

    Eaton-Rosen, F

    Z. Eaton-Rosen, F. Bragman, S. Bisdas, S. Ourselin, M. J. Cardoso, Towards safe deep learning: accurately quantifying biomarker uncer- tainty in neural network predictions, in: Medical Image Computing and Computer Assisted Intervention–MICCAI 2018: 21st International Con- fere...

    2018

  28. [36]

    Y . Gal, Z. Ghahramani, Dropout as a bayesian approximation: Repre- senting model uncertainty in deep learning, in: international conference on machine learning, PMLR, 2016, pp. 1050–1059

    2016

  29. [37]

    J. M. Hern ´andez-Lobato, R. Adams, Probabilistic backpropagation for scalable learning of bayesian neural networks, in: International confer- ence on machine learning, PMLR, 2015, pp. 1861–1869

    2015

  30. [38]

    Hernandez-Lobato, Y

    J. Hernandez-Lobato, Y . Li, M. Rowland, T. Bui, D. Hern ´andez-Lobato, R. Turner, Black-box alpha divergence minimization, in: International conference on machine learning, PMLR, 2016, pp. 1511–1520

    2016

  31. [39]

    G. Wang, W. Li, M. Aertsen, J. Deprest, S. Ourselin, T. Vercauteren, Aleatoric uncertainty estimation with test-time augmentation for medical image segmentation with convolutional neural networks, Neurocomput- ing 338 (2019) 34–45

    2019

  32. [40]

    Nixon, M

    J. Nixon, M. W. Dusenberry, L. Zhang, G. Jerfel, D. Tran, Measuring calibration in deep learning., in: CVPR workshops, V ol. 2, 2019

    2019

  33. [41]

    A. P. Dempster, Upper and lower probabilities induced by a multivalued mapping, in: Classic works of the Dempster-Shafer theory of belief func- tions, Springer, 2008, pp. 57–72

    2008

  34. [42]

    Fidon, M

    L. Fidon, M. Aertsen, F. Kofler, A. Bink, A. L. David, T. Deprest, D. Emam, F. Gu ffens, A. Jakab, G. Kasprian, et al., A dempster-shafer approach to trustworthy ai with application to fetal brain mri segmen- tation, IEEE transactions on pattern analysis and machine intelligen...

    2024

  35. [43]

    Lakshminarayanan, A

    B. Lakshminarayanan, A. Pritzel, C. Blundell, Simple and scalable pre- dictive uncertainty estimation using deep ensembles (2016)

    2016

  36. [44]

    Kiela, E

    D. Kiela, E. Grave, A. Joulin, T. Mikolov, E fficient large-scale multi- modal classification (2018)

    2018

  37. [45]

    Todisco, X

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T. Kinnunen, K. A. Lee, Asvspoof 2019: Future horizons in spoofed and fake audio detection, arXiv preprint arXiv:1904.05441 (2019)

    2019 arXiv

  38. [46]

    Yamagishi, X

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, X. Liu, K. A. Lee, T. Kinnunen, N. Evans, et al., Asvspoof 2021: accel- erating progress in spoofed and deepfake speech detection, in: ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoofing Co...

    2021

  39. [47]

    Kinnunen, H

    T. Kinnunen, H. Delgado, N. Evans, K. A. Lee, V . Vestman, A. Nautsch, M. Todisco, X. Wang, M. Sahidullah, J. Yamagishi, D. A. Reynolds, Tandem Assessment of Spoofing Countermeasures and Au- tomatic Speaker Verification: Fundamentals, IEEE /ACM Transactions on Audio, Speech, a...

    2020

  40. [48]

    G. Hua, A. B. J. Teoh, H. Zhang, Towards end-to-end synthetic speech detection, IEEE Signal Processing Letters 28 (2021) 1265–1269

    2021

  41. [49]

    X. Li, N. Li, C. Weng, X. Liu, D. Su, D. Yu, H. Meng, Replay and syn- thetic speech detection with res2net architecture, in: ICASSP 2021-2021 IEEE international conference on acoustics, speech and signal processing (ICASSP), IEEE, 2021, pp. 6354–6358

    2021

  42. [50]

    X. Ma, T. Liang, S. Zhang, S. Huang, L. He, Improved lightcnn with attention modules for asv spoofing detection, in: 2021 IEEE International Conference on Multimedia and Expo (ICME), IEEE, 2021, pp. 1–6

    2021

  43. [51]

    Z. Lei, Y . Yang, C. Liu, J. Ye, Siamese convolutional neural network us- ing gaussian probability feature for spoofing speech detection., in: Inter- speech, 2020, pp. 1116–1120

    2020

  44. [52]

    C. Wang, J. Yi, J. Tao, H. Sun, X. Chen, Z. Tian, H. Ma, C. Fan, R. Fu, Fully automated end-to-end fake audio detection, in: Proceedings of the 1st International Workshop on Deepfake Detection for Audio Multimedia, 2022, pp. 27–33

    2022

  45. [53]

    C ´aceres, R

    J. C ´aceres, R. Font, T. Grau, J. Molina, B. V . SL, The biometric vox sys- tem for the asvspoof 2021 challenge, in: Proc. ASVspoof2021 Workshop, 2021

    2021

  46. [54]

    Y . Wen, Z. Lei, Y . Yang, C. Liu, M. Ma, Multi-path gmm-mobilenet based on attack algorithms and codecs for synthetic speech and deepfake detec- 10 tion., in: INTERSPEECH, 2022, pp. 4795–4799

    2022

  47. [55]

    W. H. Kang, J. Alam, A. Fathan, Crim’s system description for the asvspoof2021 challenge, in: Proc. ASVspoof 2021 Workshop, 2021, pp. 100–106

    2021

  48. [56]

    Fathan, J

    A. Fathan, J. Alam, W. H. Kang, Mel-spectrogram image-based end-to- end audio deepfake detection under channel-mismatched conditions, in: 2022 IEEE International Conference on Multimedia and Expo (ICME), IEEE, 2022, pp. 1–6. 11

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.