Pith. sign in

REVIEW 3 major objections 5 minor 69 references

Dataset identity alone can serve as a supervisory signal to improve audio deepfake detection across heterogeneous datasets, cutting Average EER by 13.14% and Pooled EER by 5.32%.

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

2026-07-31 23:24 UTC pith:GD3FC2Z7

load-bearing objection Practical idea with a real evaluation-reporting issue: the dataset-identity trick mostly holds up on held-out data, but the 'unseen' claim overstates the aggregate. the 3 major comments →

arxiv 2607.23961 v1 pith:GD3FC2Z7 submitted 2026-07-27 eess.AS

Leveraging Gradient Reversal Loss and Multitask Learning for Datasets-Aware Audio Deepfake Detection

Mingrui Liang , Thomas Thebaud , Lukasz Wojciak , Laureano Moro Velazquez , Yishay Carmiel , Jesus Villalba Lopez , Najim Dehak This is my paper
classification eess.AS
keywords audio deepfake detectionspeech anti-spoofingdataset identitymultitask learninggradient reversal layerdomain generalizationself-supervised learningequal error rate
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 argues that the one piece of metadata every training corpus has — the identity of the dataset itself — can be used as an auxiliary training signal to make a deepfake detector generalize better across heterogeneous sources. Rather than relying on language, codec, or spoofing-method annotations, which are inconsistently available, the authors add dataset identity as supervision in two ways: a multitask branch that predicts a combined dataset×spoof/bona-fide label, and a gradient-reversal branch that adversarially removes dataset-specific information from learned features. On the 2025 Speech DeepFake Arena evaluation, the multitask version reduces Average EER from 9.484% to 8.238% (13.14% relative), and the GRL version reduces Pooled EER from 12.596% to 11.926% (5.32% relative), using a single 315.4M-parameter detector. The authors conclude that dataset identity is a simple and effective supervisory signal for improving aggregate detection performance across diverse evaluation sets.

Core claim

The central discovery is that dataset identity, the most basic piece of provenance information in any mixed-data training collection, can serve as a stand-in for richer auxiliary labels that are often missing. The paper introduces two training objectives around this cue. In multitask learning (MT), the auxiliary classifier predicts class-conditional labels of the form (dataset, spoof/bona fide), which keeps the auxiliary task aligned with the main spoof-detection objective while capturing dataset-specific variation. In gradient-reversal learning (GRL), the auxiliary classifier tries to predict the dataset identity from features, and the backbone is trained adversarially to make that predicti

What carries the argument

The class-conditional dataset label y_aux=(dataset, spoof/bona fide) for the multitask branch, and the gradient-reversal layer (GRL) for the adversarial branch. GRL is a module that passes features unchanged in the forward pass but reverses the gradient of the auxiliary dataset-classification loss during backpropagation, so the backbone learns representations from which dataset identity cannot be recovered. Both objectives are added to the main binary spoof/bona-fide cross-entropy loss with a weight λ=0.1. The backbone is an XLS-R SSL encoder with attentive layer fusion and a single SE-Res2Block ECAPA-TDNN, 315.4M parameters total, trained in two stages (frozen encoder, then full fine-tuning

Load-bearing premise

The evaluation protocol includes several datasets that were also used in training (ASVspoof 2019, ASVspoof 2024, Fake or Real, Codecfake, DFADD, LibriSeVoc, SONAR), so the aggregate EER improvements may partly come from the model memorizing training-set artifacts rather than from generalizing to genuinely unseen data.

What would settle it

Recompute Average and Pooled EER using only the truly held-out evaluation subsets (ASVspoof 2021 LA/DF, In The Wild, ADD2022 Tracks 1 and 3, ADD2023 Rounds 1 and 2). If MT no longer beats the baseline on Average EER, or GRL no longer beats the baseline on Pooled EER, on this subset, then the claimed cross-dataset generalization improvement is an artifact of the evaluation protocol. Alternatively, train the same framework with randomly shuffled or constant dataset labels; if the performance gains persist, the effect is not attributable to dataset structure.

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

If this is right

  • If dataset identity is a sufficient auxiliary signal, then any collection of mixed datasets—even unlabeled in-the-wild corpora—can be plugged into training without needing language, codec, or spoofing-method annotations.
  • The complementary gains of MT (per-dataset average) and GRL (pooled threshold behavior) suggest a unified objective combining both could improve both aggregate metrics simultaneously; the paper leaves this as future work.
  • The method achieves competitive leaderboard results with a 315.4M-parameter model, whereas the two systems above it use 500M and 1000M parameters, indicating a favorable accuracy/size trade-off.
  • Per-dataset results show large gains on several evaluation subsets (ASVspoof 2019 from 5.850% to 1.526% EER; Fake or Real from 1.428% to 0.057%) for MT, suggesting dataset-conditioned supervision can substantially improve detection in benchmark evaluations.

Where Pith is reading between the lines

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

  • The evaluation protocol mixes datasets that also appear in training (ASVspoof 2019/2024, Fake or Real, Codecfake, DFADD, LibriSeVoc, SONAR) with truly held-out subsets; the aggregate Average/Pooled numbers therefore blend in-domain and out-of-domain performance. Recomputing the metrics on the held-out subsets alone (ASV2021 LA/DF, In The Wild, ADD2022 T1/T3, ADD2023 R1/R2) would reveal whether the
  • The per-dataset held-out results are mixed: GRL improves In The Wild (5.211→3.062) and ADD2023 R2 (23.935→18.290) but degrades ASV2021 LA (10.940→13.712) and ADD23 R1 (15.978→18.389). So the out-of-domain evidence does not uniformly support the headline generalization claim.
  • If the benefit is truly driven by dataset identity as a grouping signal, the same recipe could be applied to any consistently available grouping attribute (e.g., recording device, channel, or collection batch); testing this would distinguish 'dataset identity as a useful meta-label' from 'specific dataset artifacts.'
  • A unified MT+GRL objective, which the paper names as future work, could potentially yield both low Average and low Pooled EER; this is the immediate next test.

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 / 5 minor

Summary. The paper proposes using dataset identity as an auxiliary supervisory signal for audio deepfake detection, in two forms: multitask learning with class-conditional dataset labels (dataset × spoof/bona fide) and gradient reversal layer (GRL) training that adversarially suppresses dataset identity. The system is a single 315.4M-parameter XLS-R + ECAPA-TDNN detector baseline. Experiments follow the 2025 Speech Deepfake Arena protocol across 14 evaluation subsets. The paper reports that MT reduces Average EER from 9.484% to 8.238% (13.14% relative) and GRL reduces Pooled EER from 12.596% to 11.926% (5.32% relative). Ablations compare dataset-only vs. class-conditional labels for MT and language vs. dataset labels for GRL. The central claim is that dataset identity is a simple, effective, and practical supervisory signal that improves aggregate detection performance across heterogeneous evaluation datasets, including unseen real-world data.

Significance. If the reported gains are robust, the work is a useful practical contribution: it avoids requiring auxiliary metadata (e.g., language, codec) that is often unavailable in mixed-dataset training, and it requires no increase in model size. The paper provides public model repositories, a two-stage training recipe, and a leaderboard comparison showing competitive performance at 315.4M parameters versus 500M/1B systems. The ablation on auxiliary label design is a strength: it shows that the class-conditional MT label and dataset-based GRL label each outperform simpler alternatives. The central generalization claim, however, is weakened by the evaluation protocol's training/evaluation overlap and by mixed held-out results, so the significance as a generalization result is currently not fully established.

major comments (3)
  1. [§4.1, Table 1] The evaluation protocol includes training datasets. Seven of the fourteen evaluation subsets (ASVspoof 2019, ASVspoof 2024, Fake or Real, Codecfake, DFADD, LibriSeVoc, SONAR) appear in both the Train and Eval columns. The headline Average EER (9.484→8.238) and Pooled EER (12.596→11.926) gains aggregate in-domain subsets with truly held-out subsets (ASV2021, In The Wild, ADD2022/2023). Since the abstract claims improved performance on 'unseen real-world data,' the aggregate improvements do not by themselves support that claim; they may be driven by memorizing training-set artifacts. Please report the aggregate metrics restricted to held-out subsets, or otherwise demonstrate that the gains persist when training-overlapping subsets are excluded.
  2. [§5.1, Table 2] The held-out subset evidence is mixed. GRL improves In The Wild (5.211→3.062) and ADD 2023 R2 (23.935→18.290) but degrades ASV2021 LA (10.940→13.712) and ADD 2023 R1 (15.978→18.389). MT improves most held-out subsets but degrades ASV2024-EVAL (13.263→14.371) and ADD 2023 R1 (15.978→19.967). The paper's conclusion that dataset-identity supervision improves 'cross-dataset generalization' is therefore not consistently supported by the held-out data. Please either temper the generalization claim to 'aggregate performance on the Arena benchmark' or provide a statistical test showing the held-out gains are unlikely to be due to chance.
  3. [§4.3, §5.1] No error bars, confidence intervals, or multi-seed results are reported. The key improvements are numerically small: GRL's Pooled EER improvement is 0.67% absolute, and MT's Average EER improvement is 1.246% absolute. The auxiliary loss weight λ=0.1 is tuned on 'a smaller development subset' (§4.3), which is reasonable, but without variance estimates it is unclear whether the reported differences are stable across training runs. Please add at least bootstrap confidence intervals on EER or results from multiple training seeds for the baseline, MT, and GRL systems.
minor comments (5)
  1. [Table 2] The label 'ASV2024-EV AL' appears to be a typo for 'ASV2024-EVAL'; please correct for consistency with the note.
  2. [§2.3] The text says 'current models have thousands of parameters,' but the model in this paper has 315.4M parameters; presumably 'millions' or 'hundreds of millions' is intended.
  3. [References] References [25] and [56] appear to cite the same paper ('Do compact SSL backbones matter for audio deepfake detection? A controlled study with RAPTOR'); please deduplicate or clarify the distinction.
  4. [Figure 3] The t-SNE plots are described qualitatively, but the figure is not provided in the text; if included in the final version, ensure the dataset-type and spoof-det colorings are legible and that the sampling procedure (equal bona fide/spoof per dataset) is stated in the caption.
  5. [§5.3, Table 4] The text says 'Dataset type×Spoof det achieves better results on 9 out of 14 datasets,' but the table shows it is worse on LibriSeVoc, DFADD, and ASV2024-EVAL, which is 3 of 14; the count is correct only if ties are excluded, so please clarify the counting rule.

Circularity Check

0 steps flagged

No significant circularity: the paper reports external benchmark measurements with hyperparameters tuned on a development subset, and the central claim is an empirical result rather than a derivation from its own definitions.

full rationale

The paper makes no formal derivation whose conclusion is equivalent to its premises. The method is defined operationally (Eqs. 1–16: MT auxiliary label y_MT_aux = (d, y), GRL gradient reversal), and the reported improvements in Average EER and Pooled EER are obtained by running the trained models on the 2025 Speech Deepfake Arena evaluation protocol. Hyperparameters (lambda = 0.1, AAM scale, margin, learning rates) are set on a smaller development subset, not by fitting the evaluation EERs, so the aggregate numbers are not forced by construction. The baseline is adopted from cited prior work ([54]), and the evaluation protocol is from an external benchmark ([55]); no load-bearing claim rests on a self-citation or on an imported uniqueness theorem. The fact that seven evaluation subsets also appear in the training column of Table 1 is a legitimate threat to the generalization interpretation of the aggregate gains, and the held-out subset results are mixed; however, that is an experimental-validity concern about protocol overlap, not a circularity of the kind where a fitted input is renamed as a prediction or where X is defined in terms of Y. Accordingly, no specific circular step can be exhibited, and the honest finding is no significant circularity (score 0).

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central training recipe rests on standard neural-network assumptions plus the availability of clean dataset identity and spoof labels. The only fitted quantities are conventional loss weights and AAM hyperparameters chosen on a development subset; no invented entities are introduced.

free parameters (4)
  • lambda_MT = 0.1
    Auxiliary loss weight for multitask objective (Eq. 7); set "based on preliminary experiments on a smaller development subset" (§4.3).
  • lambda_GRL = 0.1
    Auxiliary loss weight for GRL objective (Eq. 15); same tuning procedure (§4.3).
  • AAM-Softmax margin = 0.2
    Margin in the main classification loss in Stage 2 (§4.3); chosen by hand, affects learned embedding geometry.
  • AAM-Softmax scale = 32
    Scale in the main classification loss in Stage 2 (§4.3); chosen by hand.
axioms (4)
  • domain assumption Dataset identity is a consistent, available label for every training sample in the pooled corpus.
    The method's auxiliary labels yMT_aux=(d,y) and yGRL_aux=d (Eqs. 1, 11) require d to be known and meaningful; if dataset labels are noisy or inconsistent, the auxiliary tasks are ill-posed (§1, Table 1).
  • domain assumption Bona fide/spoof labels are correct and directly comparable across all pooled datasets.
    The main loss and the class-conditional auxiliary label (d,y) assume the binary spoof label y is reliable in every dataset; label noise or differing definitions would distort both tasks (§4.1).
  • standard math Gradient reversal encourages domain-invariant features that generalize to new datasets.
    GRL (Eqs. 13–16) relies on the standard adversarial-training premise that removing dataset-discriminative information preserves task-relevant cues; this is a known property but an assumption for unseen domains (§3.2).
  • domain assumption The mixed Arena evaluation protocol, including in-domain and out-of-domain sets, is a valid proxy for real-world generalization.
    The interpretation of the aggregate EER gains as "unseen real-world data" depends on this premise, but Table 1 shows many evaluation datasets also used in training.

pith-pipeline@v1.3.0-alltime-deepseek · 14454 in / 14355 out tokens · 131746 ms · 2026-07-31T23:24:06.286914+00:00 · methodology

0 comments
read the original abstract

Recent advances in speech synthesis and voice conversion, which pose threats to security and privacy, have underscored the need for deepfake detection technology. Although existing detection systems achieve strong performance on individual datasets, they often fail to generalize across diverse datasets. Prior methods for improving generalization, including data augmentation, adversarial training on auxiliary factors such as language or codec types, and Mixture-of-Experts (MoE), are limited by predefined augmentation coverage, difficulties in obtaining auxiliary factors, and substantial model complexity. In this work, we propose a practical dataset-aware framework for deepfake detection. Our method targets heterogeneous datasets for which auxiliary annotations such as language, codec, or spoofing method may not be consistently available. We therefore rely only on dataset identity as a naturally available supervisory signal for multitask (MT) and gradient reversal layer (GRL) training, allowing the model to investigate both dataset-aware multitask supervision and adversarial suppression of dataset-specific information. We conduct experiments following the 2025 Speech DeepFake Arena benchmark protocol, evaluating our model across multiple evaluation datasets and reporting aggregate performance in terms of Equal Error Rate (EER), including Average EER and Pooled EER. Compared with the baseline, MT reduces Average EER by 13.14% relatively, while GRL reduces Pooled EER by 5.32% relatively. These results demonstrate that our method can improve aggregate detection performance across heterogeneous evaluation datasets, offering a practical solution for deploying reliable deepfake detection systems on diverse and unseen real-world data.

Figures

Figures reproduced from arXiv: 2607.23961 by Jesus Villalba Lopez, Laureano Moro Velazquez, Lukasz Wojciak, Mingrui Liang, Najim Dehak, Thomas Thebaud, Yishay Carmiel.

Figure 1
Figure 1. Figure 1: Overview of the unified data construction. After ag￾gregating datasets from diverse sources, only bona fide/spoof labels and dataset identity remain consistently available across datasets. The Curated, Challenge, and In-the-Wild categories indicate the primary grouping of each dataset and are not in￾tended to be strictly mutually exclusive. become an active research field [2, 3]. Initially, researchers pri… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. The blue modules represent the baseline detection pipeline. MT consists of the blue modules and the green auxiliary branch for multitask learning, while GRL consists of the blue modules and the orange auxiliary branch. The snowflake and flame icons denote frozen and trainable modules, respectively, with the feature extractor frozen in Stage 1 and all modules trainable in… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of learned embeddings under different training strategies. Each row corresponds to one model configu￾ration: Baseline, MT, and GRL. Within each row, the two panels show the same embedding space colored by Dataset type labels and Spoof det labels, respectively. “‘ Among the open systems summarized in the table, MT achieves the third-best Average EER, while GRL achieves the third-best Poo… 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

69 extracted references · 20 linked inside Pith

  1. [1]

    Introduction With the help of powerful deep neural networks, recent ad- vances in text-to-speech (TTS) and voice conversion (VC) have significantly improved the perceptual quality of generated speech. Although this technique is widely used across domains, such as data enhancement [1], it can also be used for criminal activities, including financial fraud,...

    Pith/arXiv arXiv 2026

  2. [2]

    Hand-crafted Feature-based Audio Deepfake Detec- tion Hand-crafted features (e.g., Mel Spectrogram) are widely used across many speech-related fields

    Related Work 2.1. Hand-crafted Feature-based Audio Deepfake Detec- tion Hand-crafted features (e.g., Mel Spectrogram) are widely used across many speech-related fields. In preliminary audio deep- fake detection work, these features indeed achieved some re- markable results [16, 17]. For example, HM-Conformer [18] introduced the hierarchical pooling and mu...

  3. [3]

    Considering this, we propose two methods to explicitly allow the system to understand the special information: multitask (MT) and gradient reversal layer (GRL)

    Method To make the system robust across scenarios, the key is to enable it to learn characteristics that are irrelevant to those scenarios, such as dataset-specific ones. Considering this, we propose two methods to explicitly allow the system to understand the special information: multitask (MT) and gradient reversal layer (GRL). Our baseline builds on th...

    2015

  4. [4]

    Datasets We construct a large-scale unified corpus for training and vali- dation by combining multiple publicly available datasets

    Experiment 4.1. Datasets We construct a large-scale unified corpus for training and vali- dation by combining multiple publicly available datasets. The usage of each dataset across the training, validation, and eval- uation splits is summarized in Table 1. Specifically, for train- ing and validation, we have ASVspoof 2015, ASVspoof 2019, ASVspoof 2024, Fa...

    arXiv 2015

  5. [5]

    Results and Analysis 5.1. MT & GRL Results Table 2 presents the evaluation results of the baseline, multi- task learning (MT), and gradient reversal layer (GRL) meth- ods on the 2025 Speech Deepfake Arena benchmark. Overall, MT achieves the best Average EER, reducing the baseline from 9.484% to 8.238%, corresponding to a relative improvement of 13.14%. Th...

    arXiv 2025

  6. [6]

    Conclusion In this work, we propose a dataset-aware framework to improve audio deepfake detection across heterogeneous evaluation set- tings. Instead of relying on auxiliary annotations such as lan- guage, codec, or spoofing method, which are often unavailable or inconsistent across datasets, our method uses dataset iden- tity as a naturally available sup...

    2025

  7. [7]

    First, MT and GRL are investigated as two separate strategies

    Limitation & Future Work Despite the promising results, this work still has several lim- itations. First, MT and GRL are investigated as two separate strategies. As shown in the results, MT achieves better Aver- age EER by learning domain-aware representations, while GRL achieves better Pooled EER by encouraging domain-invariant representations. However, ...

  8. [8]

    We thank Jan Vainer from Meaning for his helpful assistance

    Acknowledgments This work was supported in part by computational resources provided by the Center for Language and Speech Processing (CLSP) at Johns Hopkins University. We thank Jan Vainer from Meaning for his helpful assistance

  9. [9]

    All research ideas, experimental design, implemen- tation, data analysis, results, and conclusions were developed and finalized by the author

    Generative AI Use Disclosure The author used generative AI tools to assist with language pol- ishing, grammar correction, writing organization, and LaTeX formatting. All research ideas, experimental design, implemen- tation, data analysis, results, and conclusions were developed and finalized by the author. The author takes full responsibility for the con...

  10. [10]

    Asr model adaptation for rare words using synthetic data generated by multiple text- to-speech systems,

    K. C. Yuen, L. Haoyang, and C. E. Siong, “Asr model adaptation for rare words using synthetic data generated by multiple text- to-speech systems,” in2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2023, pp. 1771–1778

    2023

  11. [11]

    The defender’s perspective on automatic speaker verification: An overview,

    H. Wu, J. Kang, L. Meng, H. Meng, and H.-y. Lee, “The defender’s perspective on automatic speaker verification: An overview,”arXiv preprint arXiv:2305.12804, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    V oice spoof- ing countermeasures: Taxonomy, state-of-the-art, experimental analysis of generalizability, open challenges, and the way for- ward,

    A. Khan, K. M. Malik, J. Ryan, and M. Saravanan, “V oice spoof- ing countermeasures: Taxonomy, state-of-the-art, experimental analysis of generalizability, open challenges, and the way for- ward,”arXiv preprint arXiv:2210.00417, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    A comprehensive survey with critical analysis for deepfake speech detection,

    L. Pham, P. Lam, D. Tran, H. Tang, T. Nguyen, A. Schindler, F. Skopik, A. Polonsky, and H. C. Vu, “A comprehensive survey with critical analysis for deepfake speech detection,”Computer Science Review, vol. 57, p. 100757, 2025

    2025

  14. [14]

    Improving automatic speech recogni- tion performance for low-resource languages with self-supervised models,

    J. Zhao and W.-Q. Zhang, “Improving automatic speech recogni- tion performance for low-resource languages with self-supervised models,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1227–1241, 2022

    2022

  15. [15]

    Espnet- spk: full pipeline speaker embedding toolkit with reproducible recipes, self-supervised front-ends, and off-the-shelf models,

    J.-w. Jung, W. Zhang, J. Shi, Z. Aldeneh, T. Higuchi, B.- J. Theobald, A. H. Abdelaziz, and S. Watanabe, “Espnet- spk: full pipeline speaker embedding toolkit with reproducible recipes, self-supervised front-ends, and off-the-shelf models,” arXiv preprint arXiv:2401.17230, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    A survey on speech deep- fake detection,

    M. Li, Y . Ahmadiadli, and X.-P. Zhang, “A survey on speech deep- fake detection,”ACM Computing Surveys, vol. 57, no. 7, pp. 1–38, 2025

    2025

  17. [17]

    Singing voice graph modeling for singfake detection,

    X. Chen, H. Wu, J.-S. R. Jang, and H.-y. Lee, “Singing voice graph modeling for singfake detection,”arXiv preprint arXiv:2406.03111, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Slim: Style- linguistics mismatch model for generalized audio deepfake de- tection,

    Y . Zhu, S. Koppisetti, T. Tran, and G. Bharaj, “Slim: Style- linguistics mismatch model for generalized audio deepfake de- tection,”Advances in Neural Information Processing Systems, vol. 37, pp. 67 901–67 928, 2024

    2024

  19. [19]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,”arXiv preprint arXiv:2202.12233, 2022

    Pith/arXiv arXiv 2022

  20. [20]

    A robust audio deepfake detection system via multi-view fea- ture,

    Y . Yang, H. Qin, H. Zhou, C. Wang, T. Guo, K. Han, and Y . Wang, “A robust audio deepfake detection system via multi-view fea- ture,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 13 131–13 135

    2024

  21. [21]

    V oice deepfake detec- tion using the self-supervised pre-training model hubert,

    L. Li, T. Lu, X. Ma, M. Yuan, and D. Wan, “V oice deepfake detec- tion using the self-supervised pre-training model hubert,”Applied sciences, vol. 13, no. 14, p. 8488, 2023

    2023

  22. [22]

    Ur channel-robust syn- thetic speech detection system for asvspoof 2021,

    X. Chen, Y . Zhang, G. Zhu, and Z. Duan, “Ur channel-robust syn- thetic speech detection system for asvspoof 2021,”arXiv preprint arXiv:2107.12018, 2021

    arXiv 2021

  23. [23]

    Transferring audio deepfake detection capability across lan- guages,

    Z. Ba, Q. Wen, P. Cheng, Y . Wang, F. Lin, L. Lu, and Z. Liu, “Transferring audio deepfake detection capability across lan- guages,” inProceedings of the ACM Web Conference 2023, 2023, pp. 2033–2044

    2023

  24. [24]

    An empirical study on channel effects for synthetic voice spoofing countermeasure systems,

    Y . Zhang, G. Zhu, F. Jiang, and Z. Duan, “An empirical study on channel effects for synthetic voice spoofing countermeasure systems,”arXiv preprint arXiv:2104.01320, 2021

    arXiv 2021

  25. [25]

    Vfd-net: V ocoder fingerprints detection for fake audio,

    J. Deng, Y . Ren, T. Zhang, H. Zhu, and Z. Sun, “Vfd-net: V ocoder fingerprints detection for fake audio,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2024, pp. 12 151–12 155

    2024

  26. [26]

    Synthetic speech detection based on temporal consistency and distribution of speaker features,

    Y . Zhang, Z. Li, J. Lu, W. Wang, and P. Zhang, “Synthetic speech detection based on temporal consistency and distribution of speaker features,”arXiv preprint arXiv:2309.16954, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Hm-conformer: A conformer-based audio deepfake detection system with hierarchical pooling and multi-level classification to- ken aggregation methods,

    H.-s. Shin, J. Heo, J.-h. Kim, C.-y. Lim, W. Kim, and H.-J. Yu, “Hm-conformer: A conformer-based audio deepfake detection system with hierarchical pooling and multi-level classification to- ken aggregation methods,” inICASSP 2024-2024 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 581–10 585

    2024

  28. [28]

    Conformer: Convolution- augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wuet al., “Conformer: Convolution- augmented transformer for speech recognition,”arXiv preprint arXiv:2005.08100, 2020

    Pith/arXiv arXiv 2005

  29. [29]

    General- ized spoofing detection inspired from audio generation artifacts,

    Y . Gao, T. Vuong, M. Elyasi, G. Bharaj, and R. Singh, “General- ized spoofing detection inspired from audio generation artifacts,” arXiv preprint arXiv:2104.04111, 2021

    Pith/arXiv arXiv 2021

  30. [30]

    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. Xiaoet 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

  31. [31]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  32. [32]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,”Advances in neural information processing systems, vol. 33, pp. 12 449–12 460, 2020

    2020

  33. [33]

    Xls-r: Self- supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pinoet al., “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

    Pith/arXiv arXiv 2021

  34. [35]

    Audio deepfake detection with self-supervised xls-r and sls classifier,

    Q. Zhang, S. Wen, and T. Hu, “Audio deepfake detection with self-supervised xls-r and sls classifier,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 6765– 6773

    2024

  35. [36]

    Audio deep- fake detection with self-supervised wavlm and multi-fusion atten- tive classifier,

    Y . Guo, H. Huang, X. Chen, H. Zhao, and Y . Wang, “Audio deep- fake detection with self-supervised wavlm and multi-fusion atten- tive classifier,” inICASSP 2024-2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12 702–12 706

    2024

  36. [37]

    Comprehensive layer-wise analysis of ssl models for audio deep- fake detection,

    Y . El Kheir, Y . Samih, S. Maharjan, T. Polzehl, and S. M ¨oller, “Comprehensive layer-wise analysis of ssl models for audio deep- fake detection,” inFindings of the Association for Computational Linguistics: NAACL 2025, 2025, pp. 4070–4082

    2025

  37. [38]

    Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and ag- gregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    Pith/arXiv arXiv 2005

  38. [39]

    Bicrossmamba-st: speech deepfake detection with bidirectional mamba spectro-temporal cross-attention,

    Y . E. Kheir, T. Polzehl, and S. M¨oller, “Bicrossmamba-st: speech deepfake detection with bidirectional mamba spectro-temporal cross-attention,”arXiv preprint arXiv:2505.13930, 2025

    Pith/arXiv arXiv 2025

  39. [40]

    Aasist: Audio anti-spoofing using in- tegrated spectro-temporal graph attention networks,

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using in- tegrated spectro-temporal graph attention networks,” inICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2022, pp. 6367–6371

    2022

  40. [41]

    A superb-style benchmark of self-supervised speech models for audio deepfake detection,

    H. Ali, N. S. Adupa, S. Subramani, and H. Malik, “A superb-style benchmark of self-supervised speech models for audio deepfake detection,”arXiv preprint arXiv:2603.01482, 2026

    Pith/arXiv arXiv 2026

  41. [42]

    Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license,

    M. Korvas, O. Pl ´atek, O. Duˇsek, L. ˇZilka, and F. Jurˇc´ıˇcek, “Free English and Czech telephone speech corpus shared under the CC-BY-SA 3.0 license,” inProceedings of the Eigth Interna- tional Conference on Language Resources and Evaluation (LREC 2014), 2014, p. To Appear

    2014

  42. [43]

    Musan: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech, and noise corpus,”arXiv preprint arXiv:1510.08484, 2015

    Pith/arXiv arXiv 2015

  43. [44]

    Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,

    H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” inICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6382–6386

    2022

  44. [45]

    Leveraging mixture of experts for improved speech deepfake de- tection,

    V . Negroni, D. Salvi, A. I. Mezza, P. Bestagini, and S. Tubaro, “Leveraging mixture of experts for improved speech deepfake de- tection,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  45. [46]

    Dive into moe: Diversity-enhanced reconstruction of large lan- guage models from dense into mixture-of-experts,

    Y . Feng, B. Shen, N. Gu, J. Zhao, P. Fu, Z. Lin, and W. Wang, “Dive into moe: Diversity-enhanced reconstruction of large lan- guage models from dense into mixture-of-experts,” inProceed- ings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), 2025, pp. 19 375– 19 394

    2025

  46. [47]

    Asvspoof: The automatic speaker verification spoofing and countermeasures challenge,

    Z. Wu, J. Yamagishi, T. Kinnunen, C. Hanilc ¸i, M. Sahidullah, A. Sizov, N. Evans, M. Todisco, and H. Delgado, “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

  47. [48]

    Asvspoof 2019: Future horizons in spoofed and fake audio de- tection,

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

    Pith/arXiv arXiv 2019

  48. [49]

    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, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunenet al., “Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,”arXiv preprint arXiv:2408.08739, 2024

    Pith/arXiv arXiv 2024

  49. [50]

    For: A dataset for synthetic speech detection

    R. Reimao and V . Tzerpos, “For: A dataset for synthetic speech detection.” inSpeD, 2019, pp. 1–10

    2019

  50. [51]

    Dfadd: The diffusion and flow- matching based audio deepfake dataset,

    J. Du, I.-M. Lin, I.-H. Chiu, X. Chen, H. Wu, W. Ren, Y . Tsao, H.-Y . Lee, and J.-S. R. Jang, “Dfadd: The diffusion and flow- matching based audio deepfake dataset,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 921– 928

    2024

  51. [52]

    Mlaad: The multi- language audio anti-spoofing dataset,

    N. M. M ¨uller, P. Kawa, W. H. Choong, E. Casanova, E. G ¨olge, T. M¨uller, P. Syga, P. Sperl, and K. B¨ottinger, “Mlaad: The multi- language audio anti-spoofing dataset,” in2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024, pp. 1–7

    2024

  52. [53]

    The codecfake dataset and counter- measures for the universally detection of deepfake audio,

    Y . Xie, Y . Lu, R. Fu, Z. Wen, Z. Wang, J. Tao, X. Qi, X. Wang, Y . Liu, H. Chenget al., “The codecfake dataset and counter- measures for the universally detection of deepfake audio,”IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 386–400, 2025

    2025

  53. [54]

    Spoofceleb: Speech deepfake detection and sasv in the wild,

    J.-w. Jung, Y . Wu, X. Wang, J.-H. Kim, S. Maiti, Y . Matsunaga, H.-j. Shim, J. Tian, N. Evans, J. S. Chunget al., “Spoofceleb: Speech deepfake detection and sasv in the wild,”IEEE Open Jour- nal of Signal Processing, 2025

    2025

  54. [55]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in Acoustics, Speech and Signal Processing (ICASSP), 2015 IEEE International Conference on. IEEE, 2015, pp. 5206–5210

    2015

  55. [56]

    The first voiceprivacy attacker challenge,

    N. Tomashenko, X. Miao, E. Vincent, and J. Yamagishi, “The first voiceprivacy attacker challenge,” inICASSP 2025-2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2025, pp. 1–2

    2025

  56. [57]

    Sonar: A synthetic ai-audio de- tection framework and benchmark,

    X. Li, P.-Y . Chen, and W. Wei, “Sonar: A synthetic ai-audio de- tection framework and benchmark,” 2024

    2024

  57. [58]

    Ai-synthesized voice detection using neural vocoder artifacts,

    C. Sun, S. Jia, S. Hou, and S. Lyu, “Ai-synthesized voice detection using neural vocoder artifacts,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 904–912

    2023

  58. [59]

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

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kin- nunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautschet al., “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 2507–2522, 2023

    2021

  59. [60]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?”In- terspeech, 2022

    2022

  60. [61]

    Add 2022: the first audio deep synthe- sis detection challenge,

    J. Yi, R. Fu, J. Tao, S. Nie, H. Ma, C. Wang, T. Wang, Z. Tian, Y . Bai, C. Fanet al., “Add 2022: the first audio deep synthe- sis detection challenge,” inICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9216–9220

    2022

  61. [62]

    Add 2023: the second audio deepfake detection challenge,

    J. Yi, J. Tao, R. Fu, X. Yan, C. Wang, T. Wang, C. Y . Zhang, X. Zhang, Y . Zhao, Y . Renet al., “Add 2023: the second audio deepfake detection challenge,”arXiv preprint arXiv:2305.13774, 2023

    Pith/arXiv arXiv 2023

  62. [63]

    Exploring generalization to unseen audio data for spoofing: Insights from ssl models,

    A. Kulkarni, H. M. Tran, A. Kulkarni, S. Dowerah, D. Lolive, and M. M. Doss, “Exploring generalization to unseen audio data for spoofing: Insights from ssl models,” inASVSpoof workshop 2024, 2024

    2024

  63. [64]

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

    S. Dowerah, A. Kulkarni, A. Kulkarni, H. M. Tran, J. Kalda, A. Fedorchenko, B. Fauve, D. Lolive, T. Alum¨ae, and M. M. Doss, “Speech df arena: A leaderboard for speech deepfake detection models,”IEEE Open Journal of Signal Processing, 2026

    2026

  64. [65]

    Do compact ssl backbones matter for audio deepfake detection? a controlled study with raptor,

    A. Kulkarni, S. Dowerah, A. Kulkarni, T. Alum ¨ae, and M. M. Doss, “Do compact ssl backbones matter for audio deepfake detection? a controlled study with raptor,” 2026. [Online]. Available: https://arxiv.org/abs/2603.06164

    arXiv 2026

  65. [66]

    Temporal-channel modeling in multi-head self-attention for synthetic speech detection,

    D.-T. Truong, R. Tao, T. Nguyen, H.-T. Luong, K. A. Lee, and E. S. Chng, “Temporal-channel modeling in multi-head self-attention for synthetic speech detection,”arXiv preprint arXiv:2406.17376, 2024

    Pith/arXiv arXiv 2024

  66. [67]

    Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,

    T. Liu, D.-T. Truong, R. K. Das, K. A. Lee, and H. Li, “Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,”IEEE Transactions on Information Foren- sics and Security, vol. 20, pp. 12 005–12 018, 2025

    2025

  67. [68]

    Xlsr-mamba: A dual-column bidirec- tional state space model for spoofing attack detection,

    Y . Xiao and R. K. Das, “Xlsr-mamba: A dual-column bidirec- tional state space model for spoofing attack detection,”IEEE Sig- nal Processing Letters, 2025

    2025

  68. [69]

    Improved deepfake detection using whisper features,

    P. Kawa, M. Plata, M. Czuba, P. Sygaet al., “Improved deepfake detection using whisper features,”arXiv preprint arXiv:2306.01428, 2023

    Pith/arXiv arXiv 2023

  69. [70]

    SpeechBrain: A general-purpose speech toolkit,

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y . Gao, R. D. Mori, and Y . Ben- gio, “SpeechBrain: A general-purpose speech toolkit,” 2021, arXiv:2106.04624

    Pith/arXiv arXiv 2021