Pith. sign in

REVIEW 4 major objections 6 minor 67 references

Multi-level SSL Feature Gating for Audio Deepfake Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A single detector trained only on clean English speech reports state-of-the-art error rates on several audio deepfake benchmarks by gating all layers of the XLS-R model and forcing its classifier layers to learn different artifacts.

desk verdict Solid empirical paper with a wide multilingual evaluation, but the headline SOTA claim is weakened by kernel configuration selection on the test sets. read the letter →

arxiv 2509.03409 v1 pith:BB3CAICQ submitted 2025-09-03 cs.SD cs.AIcs.MM

Hoan My Tran , Damien Lolive , Aghilas Sini , Arnaud Delhay , Pierre-François Marteau , David Guennec This is my paper
classification cs.SDcs.AIcs.MM
keywords audiodeepfakedetectionanti-spoofingself-supervisedlearningXLS-Rmulti-kernelgatedconvolutionCenteredKernelAlignmentcross-lingualgeneralizationspeechsynthesis
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

The paper tries to show that one countermeasure, trained only on the ASVspoof 2019 Logical Access English data, can detect both known and unseen synthetic speech across many languages and attack types. It replaces the usual practice of taking the last SSL layer with a gating step that aggregates all 24 hidden layers of XLS-R, passes them through stacked multi-kernel gated convolutions, and adds a Centered Kernel Alignment loss that pushes each convolution layer to encode different artifacts. The reported results are state-of-the-art EERs on 19LA (0.08%), 21DF (1.43%), and the out-of-domain In-The-Wild set (4.44%), and the model beats a strong released-checkpoint baseline on 14 of 17 multilingual evaluation sets. If true, the practical consequence is that one modest-size detector (318M parameters) can replace a collection of specialized models.

What carries the argument

The load-bearing object is the MultiConv module paired with a CKA dissimilarity loss. MultiConv splits the projected SSL representation into two halves, runs several convolutional kernels of different sizes over one half, fuses them, and multiplies elementwise with the other half as a gate, so local artifacts at different granularities can be captured. The CKA loss, computed as the average pairwise Centered Kernel Alignment between the MultiConv layers, is added to the cross-entropy objective to penalize redundancy and force the layers to learn different synthetic-speech artifacts. A second mechanism is the SwiGLU aggregation of all 24 XLS-R hidden layers, which replaces the common choice of

What would settle it

Run the identical training pipeline with each of the seven kernel configurations in Table 4 on a fresh out-of-domain benchmark (new TTS/VC systems, new languages) that was not used anywhere in the paper. If the {3, 7, 11, 15} configuration does not maintain its advantage over the other six configurations, the reported generalization of the headline model is a selection artifact rather than a fixed-model property.

Watch

Extended reading notes

Core claim

On the paper's terms, the central claim is that aggregating all XLS-R hidden states with a SwiGLU gate, classifying with MultiConv, and training with an added CKA dissimilarity loss yields a detector that is both in-domain state-of-the-art and more general across languages and unseen attacks than prior XLS-R-based classifiers. The authors report EERs of 0.08 on 19LA, 1.43 on 21DF, 4.44 on ITW, and 2.77 on 21LA (worse than some baselines there), and state that the model achieves lower EER than the strongest alternative, XLS-R+SLS, on 14 of 17 out-of-domain evaluation sets spanning Germanic, Romance, Slavic, and Sino-Tibetan languages. Ablations show each component contributes: removing the CK

Load-bearing premise

The headline generalization numbers assume the multi-kernel configuration {3, 7, 11, 15} was fixed before the evaluation sets were inspected; the paper's own table shows no single configuration wins across all datasets, so the reported edge may be selection rather than a property of one pre-specified model.

Editorial extensions

If this is right

  • If the claim holds, one detector trained on clean English data can be deployed against unseen TTS and VC attacks, reducing the need for attack-specific retraining.
  • The 21DF result (1.43% EER) suggests the method handles vocoder artifacts and lossy compression better than previous XLS-R classifiers.
  • The multilingual out-of-domain results imply cross-lingual transfer is possible from English-only training, at least for Romance and Germanic families, with a known weakness on Chinese datasets.
  • The CKA loss offers a general recipe: enforcing inter-layer diversity in a classifier improves generalization, which could transfer to other SSL front-ends.
  • The ablations indicate data augmentation is essential; without it, 21LA EER jumps from 2.77% to 8.48%, so the gains are not purely architectural.

Reading between the lines

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

  • The paper's own Table 4 shows no single kernel configuration wins everywhere; the headline configuration {3, 7, 11, 15} happens to score best on several out-of-domain sets. A fair test would fix the kernel configuration before seeing the out-of-domain test sets, otherwise part of the reported advantage may be selection rather than architecture.
  • The CKA loss operates on the MultiConv layers, not on the 24 XLS-R layers, so it cannot directly reduce redundancy among the SSL features themselves; an extension would apply layer-wise diversity to the front-end and test whether the model generalizes further.
  • The strong HABLA result (1.45% EER) despite English-only training suggests language-independent spoofing cues exist; a testable extension is to add a small amount of noisy multilingual data and see whether the Chinese-language gap closes.
  • Most baselines in Table 2 are trained and evaluated on 4-second segments while the proposed model trains on full utterances; a controlled segment-length study would clarify how much of the gain comes from architecture versus training setup.
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 speech deepfake detection system that uses all XLS-R hidden layers, aggregated via a SwiGLU gating mechanism, followed by a MultiConv back-end with multiple convolutional kernel sizes and a CKA-based diversity loss. The model is trained on ASVspoof 2019 LA and evaluated on 19LA, 21LA, 21DF, ITW, and a broad set of out-of-domain and multilingual benchmarks. The authors report state-of-the-art results on several datasets, including 0.08% EER on 19LA, 1.43% on 21DF, and 4.44% on ITW, with ablations attributing gains to the CKA loss, SwiGLU aggregation, and data augmentation.

Significance. If the headline results hold, this is a meaningful advance: a single, relatively compact detector (318M parameters) that generalizes across multiple attack types and language families would be practically valuable and would extend prior SSL-based countermeasures. The paper's strengths include a comprehensive OOD evaluation covering 17 datasets, three-run reporting for the main configuration in Table 2, per-condition heatmaps, and a promised public code release. However, the main generalization claims currently rest on a kernel configuration that appears to be selected using the same OOD test sets, and the SOTA comparison is confounded by differing utterance-length protocols. These issues must be resolved before the central claims can be accepted.

major comments (4)
  1. [§5.1, Table 4] The kernel configuration {3,7,11,15} is the row of Table 4 that wins or nearly wins on 19LA, 21DF, ITW, ADD23-R1, ADD23-R2, and HABLA, yet the same table and the accompanying text compare seven kernel configurations on these exact evaluation sets and state that "no single configuration universally outperforms others and task-specific kernel selection is needed." This creates a serious selection-on-test risk: the OOD comparisons in Tables 2 and 3 are not evidence for a single pre-specified detector unless the kernel configuration was chosen without access to the OOD labels. I ask the authors to either fix the configuration using only the 19LA development set (or another held-out set), or report all configurations against the baselines, or provide a pre-registered selection rule.
  2. [§4.3, Table 2] The SOTA comparison is not protocol-matched. The text states that most baselines are trained and evaluated on 4-second audio segments, whereas the proposed model is trained and evaluated on full utterances. Since EER is sensitive to utterance length, the claimed improvements on 21DF (1.43% vs 1.86-1.92%) and ITW (4.44% vs 5.54-6.71%) may partly reflect this protocol difference rather than architectural superiority. Please rerun the proposed model under the 4-second segment protocol used by the baselines, or show that the ranking is invariant to segment length.
  3. [Table 3 vs Table 4 (3-runs row)] Table 3 reports a single run's OOD numbers for the proposed method, but Table 4 shows substantial run-to-run variation for the same configuration, e.g., FoR 1.81/5.66/3.80, DFADD 5.68/6.60/7.96, ADD23-R1 19.26/20.28/22.48. The "14 out of 17" claim in §4.4 is based on the favorable middle run (FoR 5.66, DFADD 6.60, ADD23-R2 17.58) rather than on the averaged results reported in Table 2. This is an unfair presentation of a stochastic result. Please report mean±std or all runs in Table 3, and state explicitly which run is used for the comparison.
  4. [Eq. (11)] The LCKA formula as written sums over q=p, including the diagonal CKA(p,p), which is always 1; the normalization by M_l(M_l-1) is then inconsistent with the number of off-diagonal pairs (M_l(M_l-1)/2). While the diagonal is a constant and may not affect gradients, the formula should be corrected to q>p or otherwise clarified. This is a technical flaw in a central equation and should be fixed in revision.
minor comments (6)
  1. [§3.3] Please specify how the activation matrices S and Y in Eq. (9) are constructed for sequence features: how are the temporal frames pooled or flattened, and what is m when using a batch? This is needed for reproducibility.
  2. [Abstract / CCS Concepts] Typo: "alignement" should be "alignment".
  3. [§5.2] Typo: "SDD" should be "SSD" (synthetic speech detection).
  4. [Table 3 caption] Typo: "Sinothe Sino-Tibetanguage" should be "Sino-Tibetan".
  5. [Table 1] The row "21LA DF" is ambiguous; please separate 21LA and 21DF or clarify the combined counts.
  6. [§4.3] Please clarify the class weighting in the weighted CE loss: is the weight 0.9 applied to the bona fide class and 0.1 to the spoofed class for every batch, and how is the imbalance in the training set accounted for?

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; empirical EERs and ablations are self-contained, with only a non-circular kernel-selection caveat.

full rationale

The paper's evidence chain is empirical, not definitional. The proposed model is specified by Eqs. (1)-(13): XLS-R hidden-state stacking, SwiGLU aggregation, MultiConv gated convolutions, MHAP pooling, and LFinal = LCE + LCKA. None of the reported EER values (Tables 2-4) is a function of these equations by construction; each is a measured error rate on held-out evaluation sets after training on 19LA train and selecting by 19LA dev performance. LCKA is an explicit regularizer that minimizes pairwise CKA between MultiConv layers; claiming that it encourages diversity is a direct statement of the loss definition, not a hidden reuse of the target metric. The ablation rows (w/o LCKA, w/o SwiGLU, w/o DA) provide independent evidence for each component. The only self-citation, [48], appears in Related Work as an adjacent task-specific layer-selection study and is not used to justify the XLS-R+MultiConv architecture or the SOTA numbers. A real methodological caveat is that Section 5.1/Table 4 sweeps seven kernel configurations on the same OOD test sets used in the headline tables and states 'no single configuration universally outperforms others and task-specific kernel selection is needed,' so the reported {3,7,11,15} configuration may benefit from post-hoc selection on test EERs. This is an evaluation-protocol/selection concern, not a circular derivation: the configuration is a hyperparameter, not a fitted parameter, and the EERs would remain empirical measurements even if the selection protocol were flawed. No step in the paper reduces a predicted quantity to a fitted input or to a self-citation, so there is no significant circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central empirical claims rest mainly on benchmark data and standard SSL features. The most consequential human-chosen parameters are the MultiConv kernel configuration and the implicit CKA weight; the kernel choice is especially important because it was selected across the same OOD test sets used for the headline claims. No new physical entities are introduced.

free parameters (4)
  • MultiConv kernel sizes = {3, 7, 11, 15}
    Chosen by hand from seven candidate sets evaluated on the same OOD benchmarks (Table 4); the paper states no single configuration dominates, so the headline numbers depend on this choice.
  • CKA loss weight = 1.0 (implicit)
    The final objective LFinal = LCE + LCKA sets the regularizer weight to 1 without a stated sweep; this controls how much inter-layer diversity is enforced.
  • Projection dimension and number of MultiConv layers = U=128, M=4
    Set in Sec. 4.3 based on prior Conformer work; only layer count is partially ablated (8 and 12 layers underperform OOD).
  • Class weights in weighted CE = 0.9 bona fide / 0.1 spoofed
    Set by hand to offset class imbalance in 19LA; affects all reported EERs.
assumptions (6)
  • domain assumption EER with LLR scores is a valid metric for countermeasure performance.
    Standard evaluation protocol for ASVspoof-style benchmarks; the paper relies on it without justifying it.
  • domain assumption XLS-R's 24 hidden layers provide complementary artifact-sensitive representations when aggregated.
    The whole front-end design depends on the value of aggregating all SSL layers rather than the last layer.
  • domain assumption Linear CKA between layer activations is a valid proxy for redundancy, so minimizing it improves feature diversity.
    The CKA loss is the paper's main novelty; its usefulness assumes CKA similarity corresponds to harmful redundancy.
  • domain assumption The public datasets and their labels are used as-is with correct protocol splits.
    All conclusions are benchmark EERs; wrong splits or labels would invalidate them.
  • domain assumption RawBoost augmentation approximates real-world distortions.
    The ablation shows huge drops without augmentation, so the central results depend on RawBoost's validity.
  • ad hoc to paper Sum-aggregation in Eq. 3 assumes projected hidden states across layers are comparable.
    The paper sums projected representations over 24 layers without normalization or justification that the layer-wise projections share a common scale and meaning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-level SSL Feature Gating for Audio Deepfake Detection." pith.science (2026). https://pith.science/paper/BB3CAICQ

@misc{pith2026250903409,
  author       = {Pith},
  title        = {Pith review of: Multi-level SSL Feature Gating for Audio Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BB3CAICQ}},
  note         = {Machine review of arXiv:2509.03409}
}
read the original abstract

Recent advancements in generative AI, particularly in speech synthesis, have enabled the generation of highly natural-sounding synthetic speech that closely mimics human voices. While these innovations hold promise for applications like assistive technologies, they also pose significant risks, including misuse for fraudulent activities, identity theft, and security threats. Current research on spoofing detection countermeasures remains limited by generalization to unseen deepfake attacks and languages. To address this, we propose a gating mechanism extracting relevant feature from the speech foundation XLS-R model as a front-end feature extractor. For downstream back-end classifier, we employ Multi-kernel gated Convolution (MultiConv) to capture both local and global speech artifacts. Additionally, we introduce Centered Kernel Alignment (CKA) as a similarity metric to enforce diversity in learned features across different MultiConv layers. By integrating CKA with our gating mechanism, we hypothesize that each component helps improving the learning of distinct synthetic speech patterns. Experimental results demonstrate that our approach achieves state-of-the-art performance on in-domain benchmarks while generalizing robustly to out-of-domain datasets, including multilingual speech samples. This underscores its potential as a versatile solution for detecting evolving speech deepfake threats.

Figures

Figures reproduced from arXiv: 2509.03409 by the authors.

Figure 1
Figure 1. Overview of the proposed model. SSL features are extracted from the input waveform. Hidden states are stacked, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Top 5 models’ performance in terms of EER (%) on [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Heatmap of performance (EER %) of our system [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Heatmap of performance (EER %) of our system [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 40 canonical work pages

  1. [48]

    Hoan My Tran, David Guennec, Philippe Martin, Aghilas Sini, Damien Lolive, Arnaud Delhay, and Pierre-François Marteau. 2024. Spoofed Speech Detection with a Focus on Speaker Embedding. In Interspeech. doi:10.21437/Interspeech. 2024-481

    doi:10.21437/interspeech 2024

  2. [1]

    Mohammed Abdelmajid H

    Khalid A. Mohammed Abdelmajid H. Mansour, Gafar Zen Alabdeen Salh. 2015. Voice Recognition using Dynamic Time Warping and Mel-Frequency Cepstral Coefficients Algorithms. International Journal of Computer Applications 116, 2 (April 2015), 34–41. doi:10.5120/20312-2362

    work page doi:10.5120/20312-2362 2015

  3. [2]

    Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henretty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor We- ber. 2020. Common Voice: A Massively-Multilingual Speech Corpus. In Pro- ceedings of the Twelfth Language Resources and Evaluation Conference , Nico- letta Calzolari, Frédéric Béchet, Philippe Blache, Khal...

    work page 2020

  4. [3]

    Zhongjie Ba, Qing Wen, Peng Cheng, Yuwei Wang, Feng Lin, Li Lu, and Zhen- guang Liu. 2023. Transferring Audio Deepfake Detection Capability across Lan- guages. In Proceedings of the ACM Web Conference 2023 (Austin, TX, USA)(WWW ’23). Association for Computing Machinery, New York, NY, USA, 2033–2044. doi:10.1145/3543507.3583222

    arXiv 2023

  5. [4]

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2022. XLS-R: Self-supervised Cross- lingual Speech Representation Learning at Scale. In Interspeech. doi:10.21437/ Interspeech.2022-143

    work page 2022

  6. [5]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representa- tions. In Advances in NeurIPS, Vol. 33

    work page 2020

  7. [6]

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Xiangzhan Yu, and Furu Wei. 2022. WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing. IEEE J. STSP 16, 6 (2022). doi:10.110...

    arXiv 2022

  8. [7]

    Orchid Chetia Phukan, Gautam Kashyap, Arun Balaji Buduru, and Rajesh Sharma

Show all 67 references
  1. [8]

    Jianfeng Deng, Lianglun Cheng, and Zhuowei Wang. 2021. Attention-based BiL- STM fused CNN with gating mechanism model for Chinese long text classification. Computer Speech & Language 68 (2021), 101182. doi:10.1016/j.csl.2020.101182

    2021

  2. [9]

    Jiawei Du, I-Ming Lin, I-Hsiang Chiu, Xuanjun Chen, Haibin Wu, Wenze Ren, Yu Tsao, Hung-Yi Lee, and Jyh-Shing Roger Jang. 2024. DFADD: The Diffusion and Flow-Matching Based Audio Deepfake Dataset. In 2024 IEEE Spoken Language Technology Workshop (SLT). 921–928. doi:10.1109/SLT...

    2024

  3. [10]

    Mark J. F. Gales, Kate M. Knill, Anton Ragni, and Shakti P. Rath. 2014. Speech recognition and keyword spotting for low-resource languages: Babel project research at CUED. In 4th Workshop on Spoken Language Technologies for Under- resourced Languages, SLTU 2014, St. Petersburg...

    2014

  4. [11]

    Albert Gu and Tri Dao. 2024. Mamba: Linear-Time Sequence Modeling with Selective State Spaces. In First Conf. on Lang. Modeling

    2024

  5. [12]

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and Ruoming Pang

  6. [15]

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. HuBERT: Self- Supervised Speech Representation Learning by Masked Prediction of Hidden Units. IEEE/ACM Tr. ASLP 29 (Oct. 2021). doi:10.1109/TASLP.2021.3122291

    2021

  7. [16]

    Edward Suh

    Weizhe Hua, Yuan Zhou, Christopher M De Sa, Zhiru Zhang, and G. Edward Suh

  8. [18]

    Miquel India, Pooyan Safari, and Javier Hernando. 2019. Self Multi-Head Atten- tion for Speaker Recognition. In Interspeech. doi:10.21437/Interspeech.2019-2616

    2019 doi

  9. [19]

    Jiachen Jiang, Jinxin Zhou, and Zhihui Zhu. 2025. Tracing Representation Pro- gression: Analyzing and Enhancing Layer-Wise Similarity. In The Thirteenth International Conference on Learning Representations . https://openreview.net/ forum?id=vVxeFSR4fU

    2025

  10. [20]

    Zehui Jin, Linlong Lang, and Biao Leng. 2025. Wave-Spectrogram Cross-Modal Aggregation for Audio Deepfake Detection. In ICASSP 2025 - 2025 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP) . 1–5. doi:10.1109/ICASSP49660.2025.10890563

    2025

  11. [22]

    Yassine El Kheir, Youness Samih, Suraj Maharjan, Tim Polzehl, and Sebastian Möller. 2025. Comprehensive Layer-wise Analysis of SSL Models for Audio Deepfake Detection. arXiv preprint arXiv:2502.03559 (2025)

    2025 arXiv

  12. [23]

    Sahidullah, Héctor Delgado, Massimiliano Todisco, Nicholas Evans, Junichi Yamagishi, and Kong Aik Lee

    Tomi Kinnunen, Md. Sahidullah, Héctor Delgado, Massimiliano Todisco, Nicholas Evans, Junichi Yamagishi, and Kong Aik Lee. 2017. The ASVspoof 2017 Challenge: Assessing the Limits of Replay Spoofing Attack Detection. In Interspeech 2017. 2–6. doi:10.21437/Interspeech.2017-1111

    2017 doi

  13. [24]

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. 2019. Similarity of neural network representations revisited. InInternational conference on machine learning. PMLR, 3519–3529

    2019

  14. [25]

    Hanxiao Liu, Zihang Dai, David So, and Quoc V Le. 2021. Pay Attention to MLPs. In Advances in NeurIPS

    2021

  15. [26]

    Xuechen Liu, Xin Wang, Md Sahidullah, Jose Patino, Héctor Delgado, Tomi Kinnunen, Massimiliano Todisco, Junichi Yamagishi, Nicholas Evans, Andreas Nautsch, and Kong Aik Lee. 2023. ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild. IEEE/ACM Transactions o...

    2023

  16. [27]

    Jaime Lorenzo-Trueba, Junichi Yamagishi, Tomoki Toda, Daisuke Saito, Fernando Villavicencio, Tomi Kinnunen, and Zhenhua Ling. 2018. The Voice Conversion Challenge 2018: Promoting Development of Parallel and Nonparallel Methods. In The Speaker and Language Recognition Workshop ...

    2018 doi

  17. [28]

    Bartłomiej Marek, Piotr Kawa, and Piotr Syga. 2024. Are audio DeepFake detection models polyglots? arXiv preprint arXiv:2412.17924 (2024)

    2024 arXiv

  18. [30]

    Nicolas Müller, Pavel Czempin, Franziska Diekmann, Adam Froghyar, and Kon- stantin Böttinger. 2022. Does Audio Deepfake Detection Generalize?. In Inter- speech. doi:10.21437/Interspeech.2022-108

    2022 doi

  19. [31]

    Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger

    Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger. 2024. MLAAD: The Multi-Language Audio Anti-Spoofing Dataset. In2024 International Joint Conference on Neural Networks (IJCNN) . ...

    2024

  20. [32]

    Sailor, and Qiongqiong Wang

    Zihan Pan, Tianchi Liu, Hardik B. Sailor, and Qiongqiong Wang. 2024. Attentive Merging of Hidden Embeddings from Pre-trained Speech Model for Anti-spoofing Detection. In Interspeech. doi:10.21437/Interspeech.2024-1472

    2024 doi

  21. [33]

    Vardan Papyan, X. Y. Han, and David L. Donoho. 2020. Prevalence of neural collapse during the terminal phase of deep learning training. Pro- ceedings of the National Academy of Sciences 117, 40 (2020), 24652–24663. arXiv:https://www.pnas.org/doi/pdf/10.1073/pnas.2015509117 doi...

    2020 doi

  22. [34]

    Ankita Pasad, Ju-Chieh Chou, and Karen Livescu. 2021. Layer-Wise Analy- sis of a Self-Supervised Speech Representation Model. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) . 914–921. doi:10.1109/ ASRU51503.2021.9688093

    2021

  23. [35]

    Ankita Pasad, Bowen Shi, and Karen Livescu. 2023. Comparative Layer-Wise Analysis of Self-Supervised Speech Models. In ICASSP 2023 - 2023 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP) . 1–5. doi:10.1109/ICASSP49357.2023.10096149

    2023

  24. [36]

    Darshan Prabhu, Yifan Peng, Preethi Jyothi, and Shinji Watanabe. 2024. MULTI- CONVFORMER: Extending Conformer with Multiple Convolution Kernels. In Interspeech. doi:10.21437/Interspeech.2024-2384 MM ’25, October 27–31, 2025, Dublin, Ireland. Hoan My Tran et al

    2024 doi

  25. [37]

    Vineel Pratap, Andros Tjandra, Bowen Shi, Paden Tomasello, Arun Babu, Sayani Kundu, Ali Elkahky, Zhaoheng Ni, Apoorv Vyas, Maryam Fazel-Zarandi, Alexei Baevski, Yossi Adi, Xiaohui Zhang, Wei-Ning Hsu, Alexis Conneau, and Michael Auli. 2024. Scaling Speech Technology to 1,000+ ...

    2024

  26. [38]

    Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. MLS: A Large-Scale Multilingual Dataset for Speech Research. In Interspeech 2020. 2757–2761. doi:10.21437/Interspeech.2020-2826

    2020 doi

  27. [39]

    Ricardo Reimao and Vassilios Tzerpos. 2019. FoR: A Dataset for Synthetic Speech Detection. In 2019 International Conference on Speech Technology and Human- Computer Dialogue (SpeD). 1–10. doi:10.1109/SPED.2019.8906599

    2019

  28. [40]

    Gomez, and Antonio Peinado

    Eros Rosello, Alejandro Gomez-Alanis, Angel M. Gomez, and Antonio Peinado

  29. [41]

    Noam Shazeer. 2020. Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)

    2020 arXiv

  30. [42]

    Chengzhe Sun, Shan Jia, Shuwei Hou, and Siwei Lyu. 2023. AI-Synthesized Voice Detection Using Neural Vocoder Artifacts. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops . 904– 912

    2023

  31. [43]

    Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, and Nicholas Evans. 2022. Rawboost: A Raw Data Boosting and Augmentation Method Ap- plied to Automatic Speaker Verification Anti-Spoofing. In ICASSP. doi:10.1109/ ICASSP43922.2022.9746213

    2022

  32. [44]

    Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee weon Jung, Junichi Yamagishi, and Nicholas Evans. 2022. Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation. In The SLR Workshop (Odyssey 2022). doi:10.21437/Odyssey.2022-16

    2022 doi

  33. [45]

    Pablo Andrés Tamayo Flórez, Rubén Manrique, and Bernardo Pereira Nunes. 2023. HABLA: A Dataset of Latin American Spanish Accents for Voice Anti-spoofing. In Interspeech 2023. 1963–1967. doi:10.21437/Interspeech.2023-2272

    2023 doi

  34. [46]

    Massimiliano Todisco, Héctor Delgado, and Nicholas Evans. 2017. Constant Q cep- stral coefficients: A spoofing countermeasure for automatic speaker verification. Computer Speech & Language 45 (2017), 516–535. doi:10.1016/j.csl.2017.01.001

    2017 doi

  35. [47]

    Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Héctor Delgado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi H Kinnunen, and Kong Aik Lee. 2019. ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection. In Interspeech

    2019

  36. [49]

    Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng. 2024. Temporal-Channel Modeling in Multi-head Self-Attention for Synthetic Speech Detection. InInterspeech. doi:10.21437/Interspeech.2024-659

    2024 doi

  37. [50]

    Jörgen Valk and Tanel Alumäe. 2021. VoxLingua107: a Dataset for Spoken Language Recognition. In Proc. IEEE SLT Workshop

    2021

  38. [51]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. In Advances in NeurIPS, Vol. 30

    2017

  39. [52]

    Bor-Shiun Wang, Chien-Yi Wang, and Wei-Chen Chiu. 2024. MCPNet: An In- terpretable Classifier via Multi-Level Concept Prototypes. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 10885–10894. doi:10.1109/CVPR52733.2024.01035

    2024

  40. [53]

    Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. Vox- Populi: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation. In Procee...

    2021

  41. [54]

    Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi

    Xin Wang, Héctor Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massim- iliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi H. Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi. 2024. ASVspoof 5: crowd- sourced speech data, deepfakes, and adversarial...

    2024 doi

  42. [55]

    Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, An- dreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, Sébastien Le Maguer, Markus Becker, F...

    2020

  43. [56]

    Zhiyong Wang, Ruibo Fu, Zhengqi Wen, Jianhua Tao, Xiaopeng Wang, Yuankun Xie, Xin Qi, Shuchen Shi, Yi Lu, Yukun Liu, et al. 2024. Mixture of experts fusion for fake audio detection using frozen wav2vec 2.0.arXiv preprint arXiv:2409.11909 (2024)

    2024 arXiv

  44. [57]

    Sahidullah, and Aleksandr Sizov

    Zhizheng Wu, Tomi Kinnunen, Nicholas Evans, Junichi Yamagishi, Cemal Hanilçi, Md. Sahidullah, and Aleksandr Sizov. 2015. ASVspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge. In Interspeech 2015. 2037–2041. doi:10.21437/Interspeech.2015-462

    2015 doi

  45. [58]

    Yang Xiao and Rohan Kumar Das. 2025. XLSR-Mamba: A Dual-Column Bidirec- tional State Space Model for Spoofing Attack Detection. IEEE Signal Processing Letters 32 (2025), 1276–1280. doi:10.1109/LSP.2025.3547861

    2025

  46. [59]

    Junichi Yamagishi, Christophe Veaux, and Kirsten MacDonald. 2019. CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit. https://api.semanticscholar.org/CorpusID:213060286

    2019

  47. [60]

    Junichi Yamagishi, Xin Wang, Massimiliano Todisco, Md Sahidullah, Jose Patino, Andreas Nautsch, Xuechen Liu, Kong Aik Lee, Tomi Kinnunen, Nicholas Evans, and Héctor Delgado. 2021. ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection. In 2021 Edition of...

    2021 doi

  48. [61]

    Jiangyan Yi, Jianhua Tao, Ruibo Fu, Xinrui Yan, Chenglong Wang, Tao Wang, Chu Yuan Zhang, Xiaohui Zhang, Yan Zhao, Yong Ren, Le Xu, Junzuo Zhou, Hao Gu, Zhengqi Wen, Shan Liang, Zheng Lian, Shuai Nie, and Haizhou Li. 2023. ADD 2023: the Second Audio Deepfake Detection Challeng...

    2023

  49. [62]

    Zhao Yi, Wen-Chin Huang, Xiaohai Tian, Junichi Yamagishi, Rohan Kumar Das, Tomi Kinnunen, Zhen-Hua Ling, and Tomoki Toda. 2020. Voice Conversion Challenge 2020 — Intra-lingual semi-parallel and cross-lingual voice conversion —. In Joint Workshop for the Blizzard Challenge and ...

    2020

  50. [63]

    What Spoofed When

    Lin Zhang, Xin Wang, Erica Cooper, Mireia Diez, Federico Landini, Nicholas Evans, and Junichi Yamagishi. 2024. Spoof Diarization: "What Spoofed When" in Partially Spoofed Audio. In Interspeech. doi:10.21437/Interspeech.2024-1365

    2024 doi

  51. [64]

    Lin Zhang, Xin Wang, Erica Cooper, Nicholas Evans, and Junichi Yamagishi

  52. [65]

    Qishan Zhang, Shuangbing Wen, and Tao Hu. 2024. Audio Deepfake Detection with Self-Supervised XLS-R and SLS Classifier. In ACM Multimedia 2024

    2024

  53. [66]

    Yuxiang Zhang, Jingze Lu, Zengqiang Shang, Wenchao Wang, and Pengyuan Zhang. 2024. Improving Short Utterance Anti-Spoofing with Aasist2. In ICASSP. doi:10.1109/ICASSP48485.2024.10448049

    2024

  54. [67]

    doi:10.21437/VCCBC.2020-14

    80–98. doi:10.21437/VCCBC.2020-14

    2020 doi

  55. [70]

    In Interspeech

    Range-Based Equal Error Rate for Spoof Localization. In Interspeech. doi:10. 21437/Interspeech.2023-1214

    2023

  56. [2019]

    In Advances in Neural Information Pro- cessing Systems, H

    Channel Gating Neural Networks. In Advances in Neural Information Pro- cessing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett (Eds.), Vol. 32. Curran Associates, Inc. https://proceedings.neurips. cc/paper_files/paper/2019/file/68b1fb...

    2019

  57. [2020]

    In Interspeech 2020

    Conformer: Convolution-augmented Transformer for Speech Recognition. In Interspeech 2020. 5036–5040. doi:10.21437/Interspeech.2020-3015

    2020 doi

  58. [2023]

    In Interspeech

    A conformer-based classifier for variable-length utterance processing in anti-spoofing. In Interspeech. doi:10.21437/Interspeech.2023-1820

    2023 doi

  59. [2024]

    In Findings of the Association for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and Steven Bethard (Eds.)

    Heterogeneity over Homogeneity: Investigating Multilingual Speech Pre- Trained Models for Detecting Audio Deepfake. In Findings of the Association for Computational Linguistics: NAACL 2024, Kevin Duh, Helena Gomez, and Steven Bethard (Eds.). Association for Computational Lingu...

    2024 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.