Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Speech DF Arena benchmarks 15 audio deepfake detectors on 14 datasets and finds that few generalize beyond their training domain.

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 ↗

Speech DF Arena standardizes audio deepfake detection benchmarking across 14 datasets and 15 systems, showing that most open-source detectors have high error rates on out-of-domain attacks.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful benchmark with a shaky 'first' claim and several numerical inconsistencies; the central message about poor cross-domain generalization still holds. the 3 major comments →

arxiv 2509.02859 v1 pith:6S7ERH7V submitted 2025-09-02 cs.SD cs.CLeess.AS

Speech DF Arena: A Leaderboard for Speech DeepFake Detection Models

Sandipana Dowerah , Atharva Kulkarni , Ajinkya Kulkarni , Hoan My Tran , Joonas Kalda , Artem Fedorchenko , Benoit Fauve , Damien Lolive
show 2 more authors
Tanel Alum\"ae Matthew Magimai Doss
This is my paper
classification cs.SD cs.CLeess.AS
keywords audio deepfake detectionanti-spoofingbenchmarkleaderboardequal error ratecross-domain generalizationself-supervised learningspeech synthesis attacks
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.

The reading

Speech DF Arena tries to become the first comprehensive benchmark that lets audio deepfake detectors be compared fairly across many datasets and attack types. It evaluates 15 systems, 12 open-source and 3 commercial, on 14 evaluation sets using a single toolkit and shared metrics. The paper's central finding is that strong within-dataset results do not imply robustness: many detectors post high equal-error rates on out-of-domain data, and no single system leads everywhere. A sensitivity study adds that even light noise or reverberation degrades top systems sharply. If this holds, the community gains a trustworthy cross-domain ranking instead of relying on single-dataset scores.

Core claim

Speech DF Arena claims to be the first comprehensive benchmark for audio deepfake detection. It evaluates 12 open-source and 3 proprietary detectors on 14 evaluation sets with one protocol and shared metrics. The central empirical claim is that detection quality is highly attack- and dataset-dependent: no system ranks first everywhere, and out-of-domain EERs are often far worse than within-dataset scores. The paper also reports that reverberation degrades even top systems by roughly half, and that pooling scores across datasets raises EER for many systems, indicating weak global-threshold generalization.

What carries the argument

The mechanism carrying the argument is the standardized evaluation protocol of the Speech DF Arena: all audio is resampled to 16 kHz, most systems receive a fixed 4-second trimmed input, batch size is 64, and the same scripts compute EER, pooled EER, accuracy, and F1 for every system on every dataset. This uniformity converts scattered per-paper numbers into one comparable table, making the out-of-domain failure pattern visible.

Load-bearing premise

The load-bearing premise is that the uniform preprocessing (16 kHz resampling and a fixed 4-second input for most systems) is neutral across detectors and that the third-party dataset labels are trustworthy ground truth; if either fails, the EER ranking reflects artifacts rather than capability.

What would settle it

Re-score all 15 systems on the same evaluation files under the standard 4-second trimmed protocol and under full-length audio, plus 2-second and 8-second cuts; if the average-EER ordering changes materially, the benchmark's cross-system ranking is an artifact of the trimming rule.

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

If this is right

  • Within-dataset EERs from published papers are not reliable indicators of deployment readiness: several systems that look strong on their home dataset exceed 20–50% EER on other benchmark sets.
  • No single architecture class dominates; ranking depends on the attack type, so choosing a detector requires knowing which deepfake generation methods are most likely.
  • For several systems, pooled EER is noticeably higher than average EER, meaning a detector with a single global threshold will misclassify more in mixed traffic than per-dataset tuning suggests.
  • Added reverberation is the most damaging simple distortion, raising EER by roughly half across the tested systems, so acoustic robustness should be part of evaluation.

Where Pith is reading between the lines

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

  • If the trimming step is not neutral, the ranking could shift; testing the same systems with variable-length inputs would reveal whether the leaderboard reflects detector quality or protocol artifacts.
  • The correlation results suggest that LibriSeVoc and ASVspoof2024 could serve as cheap screening sets; a two-dataset prescreen would catch most of the ordering signal before running the full 14-set battery.
  • A live, unlabeled evaluation mode—as the paper plans—would let new detectors be checked against freshly generated attacks without contamination, turning the leaderboard from a snapshot into a continuous robustness monitor.
Share X Bluesky LinkedIn Reddit HN

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. This paper introduces Speech DF Arena, a benchmark and leaderboard for speech deepfake detection. The authors evaluate 12 open-source and 3 proprietary detectors on 14 evaluation conditions from ASVspoof 2019/2021/2024, ADD 2022/2023, CodecFake, LibriSeVoc, SONAR, FoR, DFADD and In-the-wild, using EER, pooled EER, accuracy and F1. A toolkit is released on GitHub and the leaderboard on HuggingFace. The main empirical findings are that no single system dominates, that raw-waveform models RawNet-2 and RawGAT-ST generalize poorly, and that SSL-based mid-sized models such as XLSR+SLS have the best average/pooled EER among open-source systems; proprietary Whispeak ranks first overall. The paper also includes a correlation analysis of datasets and a MUSAN/RIR sensitivity study on three systems.

Significance. If the numbers are trustworthy, this is a useful community resource: it provides a reproducible cross-domain evaluation matrix, a toolkit, and an open leaderboard that includes both academic and industrial detectors. The qualitative result that out-of-domain EERs are often dramatically higher than within-dataset EERs (e.g., AASIST 51.05% on CodecFake) is important and likely correct. The sensitivity study is a valuable addition. However, the validity of the leaderboard hinges on the neutrality of the uniform preprocessing and on the internal consistency of reported scores; both currently require additional work.

major comments (3)
  1. [§3.1 and Table 2] The headline ranking is not protected against preprocessing artifacts. All audio is resampled to 16 kHz; WavLM/HuBERT/Wav2Vec2 systems are padded to the longest batch item, while 'the remaining systems operate on a fixed 4s input length, so input signals are trimmed accordingly.' RawNet-2, RawGAT-ST, and AASIST are raw-waveform/graph models whose native inference often consumes variable-length or longer inputs; trimming to 4 s is a protocol choice that may disproportionately lower their scores, exactly the systems the paper concludes generalize poorly. The paper provides no sensitivity analysis (e.g., 4 s vs 8 s vs full-length, with and without resampling). Without such an ablation, the Table 2 ordering cannot be attributed to detector capability rather than protocol.
  2. [§4.1 and Tables 2/3] The claim that XLSR+SLS is 'the most effective open-source system... achieving the lowest EER except the ASVspoof series' is not supported by Table 2. XLSR-Mamba has lower EER on In-the-wild (6.70 vs 7.45) and CodecFake (14.40 vs 33.43), and Nes2NetX is better on ASVspoof 2019 and 2021DF. XLSR+SLS only has the lowest average (13.84%) and pooled (15.68%) EER. Please correct the per-dataset wording and, more importantly, add uncertainty or statistical comparison for the close average gap (13.84 vs 14.21 for XLSR-Mamba) so the ranking is not over-interpreted.
  3. [§4.5 and Tables 2/4] The LibriSeVoc baselines are inconsistent across the manuscript: XLSR+SLS original EER is 1.72 in Table 2, 1.96 in Table 4, and 1.86 in the text; TCM is 2.27 vs 2.34; WavLM-ECAPA is 32.87 vs 31.88. Additionally, RawGAT-ST In-the-wild EER is 52.53 in Table 2 but 52.54 in the text. For a measurement/benchmark paper, these discrepancies undermine the reproducibility claim. Please reconcile to a single canonical table and state the exact evaluation split and metadata used for each dataset.
minor comments (5)
  1. [Abstract/Introduction] The claim of being 'the first comprehensive benchmark' should be qualified and discussed more carefully against existing works such as VoiceWukong [20] and CtrSVDD [21], which also benchmark multiple detection systems.
  2. [Table 2 header] The compact column headers 'ADD22 T1 T3 R1 R2' are ambiguous. Spell out that T1/T3 are ADD 2022 tracks and R1/R2 are ADD 2023 rounds. Use consistent dataset names ('LibriSeVoc' vs 'LibriSV').
  3. [§3.2/§2.2] The paper states that accuracy and F1 are evaluated, but these results are only mentioned as being on the HuggingFace page. Report them in the paper or clearly designate them as supplementary material.
  4. [Table 1 vs Tables 2/3] There are naming inconsistencies (e.g., 'AASIST-Large' vs 'AASIST', 'Whisper-DF-MesoNet-MFCC' vs 'Whisper Mesonet'). Standardize names across tables.
  5. [Figure 2] The AUC analysis covers only 5 systems and 5 datasets; state how these were selected and, if feasible, show variability or confidence intervals.

Circularity Check

1 steps flagged

Benchmark measurements are self-contained; only the auxiliary dataset-representativeness correlation is self-referential because the average EER includes each dataset being correlated.

specific steps
  1. self definitional [Section 3.2 and Section 4.3 (correlation analysis)]
    "We conducted the correlation analysis between different datasets and the average EER (computed across all datasets). ... We analyze the correlation of different datasets with the average EER to investigate which dataset is the best representative of the average scores."

    The 'average EER' used as the target is the mean of per-dataset EERs across the same 14 datasets. Each dataset's EER vector therefore appears on both sides of the correlation: the dataset is one of the components of the average it is being correlated with. This inflates the correlation for every dataset by construction, so the inference that LibriSeVoc and ASVspoof-2024 are 'representative' datasets is partly a self-correlation artifact rather than an independent measure of agreement. This affects only the auxiliary dataset-selection analysis in Section 4.3, not the direct leaderboard measurements in Table 2, which are computed from fixed evaluation sets with no fitted parameters.

full rationale

This is a measurement/benchmark paper, not a derivation with fitted parameters. The core results in Table 2 are EERs of pre-trained systems on fixed external evaluation sets; there is no fitted input renamed as a prediction, no load-bearing self-citation, and no ansatz smuggled in via citation. The only circular element is the correlation analysis in Section 4.3, where each dataset's EER is correlated with the average EER that includes that same dataset, making the reported 'representative dataset' correlations partially self-referential. The leaderboard ranking itself is not affected by this artifact, and the three evaluated systems from the authors' own prior work rank near the bottom, so the benchmark is not self-serving. Overall circularity is minor and localized to an auxiliary analysis.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The benchmark's claims rest on third-party dataset labels as ground truth, on released checkpoints being representative of each detector, and on the uniform preprocessing being neutral across systems. There are no scientific free parameters in the fitting sense; the 4 s trim, the 16 kHz resampling, the MUSAN SNR bands, and the pooled-EER global threshold are hand-chosen protocol values that affect the numbers. No invented entities are introduced; the leaderboard and toolkit are software artifacts, not entities postulated to explain observations.

free parameters (3)
  • Fixed input length (4 s) = 4 seconds
    All non-padding systems trim inputs to 4 s (Section 3.1); a hand-chosen protocol value that can change EER for systems trained on variable-length audio.
  • MUSAN augmentation SNR ranges = noise [0,15] dB, speech [13,20] dB, music [5,15] dB
    Hand-chosen augmentation levels for the sensitivity study (Section 4.2); the resulting EERs depend on these bands.
  • Pooled EER global threshold = derived from the pooled score distribution
    Pooled EER aggregates all scores and sets one global threshold (Section 3.2); rankings depend on this pooling choice and on the relative sizes of the datasets.
axioms (3)
  • domain assumption Dataset labels (bona fide vs spoofed) are correct ground truth
    The benchmark scores all systems against the label files of 14 third-party datasets (Section 2.1); any label noise or protocol mismatch propagates into every EER in Table 2.
  • domain assumption Official pretrained weights represent each system at SOTA performance
    Section 3.1 uses weights 'as released in their original works'; no retraining or calibration is done, so results reflect the specific released checkpoints, which may differ from the papers' reported best checkpoints.
  • domain assumption Out-of-domain status of the proprietary systems is meaningful
    Training data for Whispeak, Syntra.io, and Resemble AI is undisclosed (Section 2.2), so 'in-domain vs out-of-domain' cannot be verified for 3 of the 15 systems.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Speech DF Arena: A Leaderboard for Speech DeepFake Detection Models." pith.science (2026). https://pith.science/paper/6S7ERH7V

@misc{pith2026250902859,
  author       = {Pith},
  title        = {Pith review of: Speech DF Arena: A Leaderboard for Speech DeepFake Detection Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6S7ERH7V}},
  note         = {Machine review of arXiv:2509.02859}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Parallel to the development of advanced deepfake audio generation, audio deepfake detection has also seen significant progress. However, a standardized and comprehensive benchmark is still missing. To address this, we introduce Speech DeepFake (DF) Arena, the first comprehensive benchmark for audio deepfake detection. Speech DF Arena provides a toolkit to uniformly evaluate detection systems, currently across 14 diverse datasets and attack scenarios, standardized evaluation metrics and protocols for reproducibility and transparency. It also includes a leaderboard to compare and rank the systems to help researchers and developers enhance their reliability and robustness. We include 14 evaluation sets, 12 state-of-the-art open-source and 3 proprietary detection systems. Our study presents many systems exhibiting high EER in out-of-domain scenarios, highlighting the need for extensive cross-domain evaluation. The leaderboard is hosted on Huggingface1 and a toolkit for reproducing results across the listed datasets is available on GitHub.

Figures

Figures reproduced from arXiv: 2509.02859 by Ajinkya Kulkarni, Artem Fedorchenko, Atharva Kulkarni, Benoit Fauve, Damien Lolive, Hoan My Tran, Joonas Kalda, Matthew Magimai Doss, Sandipana Dowerah, Tanel Alum\"ae.

Figure 1
Figure 1. Figure 1: Correlation analysis between different datasets using heatmap; Dist.corr refers to Distance correlation, and CCC refers to Concordance correlation coefficient. sized transformer-based models, XLSR+SLS (340M) achieves the best balance with 13.84% average EER and 15.68% pooled EER, while XLSR Mamba (319M) shows similar average EER (14.21%) but higher pooled EER (20.12%), indicating less ro￾bust generalizatio… view at source ↗
Figure 2
Figure 2. Figure 2: AUC analysis of 5 systems on ASVspoof 24, 19, 21 LA & DF, LibriSeVoc and ADD 22 Track 1. TCM and XLSR SLS, using the XLSR backbone and being the largest system, show the best generalization over other systems [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Interface of Speech DF Arena hosted on Huggingface. out-of-distribution attacks. By testing across multiple datasets, the Speech DF Arena helps assess the generalizability and ro￾bustness of these models in real-world scenarios. We hope that our benchmarking platform fosters collaboration by creating a transparent and competitive environment, driving innovation and advancements in audio deepfake detection.… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. MixFake: Benchmarking and Enhancing Audio Deepfake Detection in Diverse Real-world Mixed Audio

    cs.SD 2026-05 unverdicted novelty 7.0

    MixFake is a new benchmark for mixed-authenticity audio and a multi-stream prompt tuning method achieves 0.95% EER foreground and 7.72% absolute gain in complex background deepfake detection.

  2. A General Model for Deepfake Speech Detection: Diverse Bonafide Resources or Diverse AI-Based Generators

    cs.SD 2026-03 unverdicted novelty 4.0

    Balancing diverse bonafide resources and AI generators in training data is the key to building general deepfake speech detection models.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages · cited by 2 Pith papers · 4 internal anchors

  1. [1]

    Introduction As AI-powered audio deepfakes continue to evolve, distinguish- ing between genuine and synthetic speech is becoming more challenging. The technological advancement achieved by the current text-to-speech (TTS) [1, 2, 3, 4] and voice conversion (VC) [5, 6] systems to obtain realism has also given rise to gen- erating audio deepfakes. Deepfakes ...

    Pith/arXiv arXiv 2025

  2. [2]

    Datasets Speech DF Arena in Phase I comprises 14 different widely rec- ognized and extensively used evaluation datasets that encom- pass common deepfake attacks

    Speech DF Arena: a unified benchmark 2.1. Datasets Speech DF Arena in Phase I comprises 14 different widely rec- ognized and extensively used evaluation datasets that encom- pass common deepfake attacks. We use the ASVspoof series (2019, 2021 and 2024)[13, 14, 30]. ASVspoof 2019 is primar- ily built for speaker verification and includes generated speech. ...

    work page 2019

  3. [3]

    Experimental setup We use readily available pre-trained weights for the open-source models, as released in their original works

    Experimentation 3.1. Experimental setup We use readily available pre-trained weights for the open-source models, as released in their original works. All selected systems are trained on the ASVspoof 2019 dataset, except for the Whisper MesoNet system, which is trained on a subset of the ASVspoof 2021 DF dataset. For benchmarking, we use only the evaluatio...

    work page 2019

  4. [4]

    Results and analysis This section presents the performance obtained by the open- source SOTA and proprietary detection systems across the datasets. 4.1. Performance of SOTA open-source systems The second row of Table 2 presents the EER of the open-source systems across various datasets. We have evaluated all the systems on 14 datasets, ensuring that model...

    work page

  5. [5]

    XLSR+SLS emerges as the most effective open-source system among all, achieving the lowest EER except the ASVspoof se- ries

    One of the key observations from Table 2 is that not a sin- gle model consistently outperforms others across all datasets. XLSR+SLS emerges as the most effective open-source system among all, achieving the lowest EER except the ASVspoof se- ries. Nes2NetX achieves the best performance in ASVspoof 2019 with an EER of 0.12% and 1.49% in ASVspoof2021DF, even...

    work page 2019

  6. [6]

    To facilitate model comparison and selection, we introduced a publicly available leaderboard on Hugging- Face, making it accessible to everyone

    Conclusion and future work We developed Speech DF Arena, a unified and comprehensive platform for benchmarking audio deepfake detection models, en- suring consistent, trustworthy, and transparent evaluation across diverse datasets. To facilitate model comparison and selection, we introduced a publicly available leaderboard on Hugging- Face, making it acce...

    work page

  7. [7]

    219726), and the Innosuisse flagship project Inclusive Information and Communication Technologies (IICT) (grant agreement no

    Acknowledgement This work was partially supported by the Swiss National Science Foundation project Pathological Speech Synthesis (PaSS)” (grant agreement no. 219726), and the Innosuisse flagship project Inclusive Information and Communication Technologies (IICT) (grant agreement no. PFFS-21-47). This work was partially supported by the Estonian Centre of ...

    work page

  8. [8]

    Towards controllable speech synthesis in the era of large language models: A survey,

    T. Xie, Y . Rong, P. Zhang, and L. Liu, “Towards controllable speech synthesis in the era of large language models: A survey,” arXiv preprint: arXiv:2412.06602, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Deepfakes detection techniques using deep learn- ing: A survey,

    A. M. Almars, “Deepfakes detection techniques using deep learn- ing: A survey,” Journal of Computer and Communications, 2021

    work page 2021

  10. [10]

    Recent advances in speech language models: A survey,

    W. Cui, D. Yu, X. Jiao, Z. Meng, G. Zhang, Q. Wang, Y . Guo, and I. King, “Recent advances in speech language models: A survey,” arXiv preprint: arXiv:2410.03751, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    knn retrieval for simple and effective zero-shot multi-speaker text-to- speech,

    K. El Hajal, A. Kulkarni, E. Hermann, and M. Magimai-Doss, “knn retrieval for simple and effective zero-shot multi-speaker text-to- speech,” in Proceedings of the Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL). ACL, 2025

    work page 2025

  12. [12]

    How far are we from robust voice conversion: A survey,

    T.-H. Huang, J.-H. Lin, C.-Y . Huang, and H.-Y . Lee, “How far are we from robust voice conversion: A survey,” inIEEE SLT, 2020

    work page 2020

  13. [13]

    An overview of voice conversion and its challenges: From statistical modeling to deep learning,

    S. K. B. Sisman, J. Yamagishi and H. Li, “An overview of voice conversion and its challenges: From statistical modeling to deep learning,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2020

    work page 2020

  14. [14]

    Exploring generalization to unseen audio data for spoof- ing: Insights from ssl models,

    A. Kulkarni, H. Tran, A. Kulkarni, S. Dowerah, D. Lolive, and M. Doss, “Exploring generalization to unseen audio data for spoof- ing: Insights from ssl models,” in ASVspoof 2024 Workshop, 2024

    work page 2024

  15. [15]

    Children’s V oice Privacy: First Steps and Emerging Challenges,

    A. Kulkarni and F. Teixeira and E. Hermann and T. Rolland and I. Trancoso and M. Magimai Doss, “Children’s V oice Privacy: First Steps and Emerging Challenges,” in Interspeech 2025, 2025, pp. 2810–2814

    work page 2025

  16. [16]

    Warning: Humans cannot reliably detect speech deepfakes,

    K. T. Mai, S. Bray, T. Davies, and L. D. Griffin, “Warning: Humans cannot reliably detect speech deepfakes,”PLOS ONE, vol. 18, no. 8, 2023

    work page 2023

  17. [17]

    Audio deepfake detection: A survey,

    J. Yi, C. Wang, J. Tao, X. Zhang, C. Y . Zhang, and Y . Zhao, “Audio deepfake detection: A survey,” 2023

    work page 2023

  18. [18]

    Towards Robust Audio Deepfake Detection: A Evolving Benchmark for Continual Learning

    X. Zhang, J. Yi, and J. Tao, “Towards robust audio deepfake detec- tion: A evolving benchmark for continual learning,”arXiv preprint: arXiv:2405.08596, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  19. [19]

    Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion,

    A. Kulkarni and S. Dowerah and T. Alumäe and M. Magimai Doss, “Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion,” in Interspeech 2025, 2025

    work page 2025

  20. [20]

    Asvspoof 2019: Future horizons in spoofed and fake audio detec- tion,

    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 detec- tion,” 2019

    work page 2019

  21. [21]

    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, and A. Nautsch, “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,” in ICASSP, 2023

    work page 2021

  22. [22]

    Asvspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detection using crowd- sourced speech,

    X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunen, N. Evans, K. A. Lee, J. Yamagishi, M. Jeong, G. Zhu, Y . Zang, Y . Zhang, S. Maiti, F. Lux, N. Müller, W. Zhang, C. Sun, S. Hou, S. Lyu, S. L. Maguer, C. Gong, H. Guo, L. Chen, and V . Singh, “Asvspoof 5: Design, collection and validation of...

    Pith/arXiv arXiv 2025

  23. [23]

    Add 2022: The first audio deep synthesis 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 synthesis detection challenge,” in ICASSP, 2022

    work page 2022

  24. [24]

    End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detec- tion,

    H. Tak, J. weon Jung, J. Patino, M. Kamble, M. Todisco, and N. Evans, “End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detec- tion,” in ASVSpoof 2021 Workshop, 2021

    work page 2021

  25. [25]

    The asvspoof 2017 challenge: As- sessing the limits of replay spoofing attack detection,

    T. Kinnunen, M. Sahidullah, H. Delgado, M. Todisco, N. Evans, J. Yamagishi, and K.-A. Lee, “The asvspoof 2017 challenge: As- sessing the limits of replay spoofing attack detection,” in Inter- speech, 2017

    work page 2017

  26. [26]

    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

  27. [27]

    V oicewukong: Benchmarking deepfake voice detection,

    Z. Yan, Y . Zhao, and H. Wang, “V oicewukong: Benchmarking deepfake voice detection,” arXiv preprint: arXiv:2409.06348 , 2024

    Pith/arXiv arXiv 2024

  28. [28]

    CtrSVDD: A Benchmark Dataset and Baseline Analysis for Controlled Singing Voice Deepfake Detection

    Y . Zang, J. Shi, Y . Zhang, R. Yamamoto, J. Han, Y . Tang, S. Xu, W. Zhao, J. Guo, T. Toda, and Z. Duan, “Ctrsvdd: A benchmark dataset and baseline analysis for controlled singing voice deepfake detection,” arXiv preprint: arXiv:2406.02438, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

    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,” in ACM Multimedia, 2024

    work page 2024

  30. [30]

    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,” in Interspeech, 2024

    work page 2024

  31. [31]

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

    Y . Xiao and R. K. Das, “Xlsr-mamba: A dual-column bidirectional state space model for spoofing attack detection,” IEEE Signal Processing Letters, vol. 32, pp. 1276–1280, 2025

    work page 2025

  32. [32]

    Im- proved deepfake detection using whisper features,

    P. Kawa, M. Plata, M. Czuba, P. Szyma ´nski, and P. Syga, “Im- proved deepfake detection using whisper features,” in Interspeech, 2023

    work page 2023

  33. [33]

    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

  34. [34]

    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,” arXiv preprint arXiv:2504.05657, 2025

    arXiv 2025

  35. [35]

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

    J. weon Jung, H.-S. Heo, H. Tak, H. jin Shim, J. S. Chung, B.-J. Lee, H. jin Yu, and N. W. D. Evans, “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” in ICASSP, 2022

    work page 2022

  36. [36]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inICASSP, 2021

    work page 2021

  37. [37]

    Asvspoof 5: Crowdsourced data, deepfakes and adversarial attacks at scale,

    X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunenet al., “Asvspoof 5: Crowdsourced data, deepfakes and adversarial attacks at scale,” in ASVspoof 2024 Workshop, 2024

    work page 2024

  38. [38]

    The codecfake dataset and countermea- sures 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 countermea- sures for the universally detection of deepfake audio,” in ICASSP, 2024

    work page 2024

  39. [39]

    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,” in CVPR, 2023

    work page 2023

  40. [40]

    Where are we in audio deepfake detection? A systematic analysis over generative and detection models

    X. Li, P.-Y . Chen, and W. Wei, “Sonar: A synthetic ai-audio detec- tion framework and benchmark,”arXiv preprint: arXiv2410.04324, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    For: A dataset for synthetic speech detection,

    R. Reimao and V . Tzerpos, “For: A dataset for synthetic speech detection,” in International Conference on Speech Technology and Human-Computer Dialogue (SpeD), 2019

    work page 2019

  42. [42]

    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. yi Lee, and J.-S. R. Jang, “Dfadd: The diffusion and flow- matching based audio deepfake dataset,” in IEEE SLT, 2024

    work page 2024

  43. [43]

    Does audio deepfake detection generalize?

    N. M. Müller, P. Czempin, F. Dieckmann, A. Froghyar, and K. Böt- tinger, “Does audio deepfake detection generalize?” Interspeech, 2022

    work page 2022

  44. [44]

    On Correlation Coefficients

    A. Stepanov, “On correlation coefficients,” arXiv preprint arXiv:2405.16469, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  45. [45]

    A review of feature selection methods based on mutual information,

    J. R. Vergara and P. A. Estévez, “A review of feature selection methods based on mutual information,” Neural Computing and Applications, vol. 24, no. 1, p. 175–186, Mar. 2013

    work page 2013

  46. [46]

    A concordance correlation coefficient to evaluate reproducibility

    L. I.-K. Lin, “A concordance correlation coefficient to evaluate reproducibility.”Biometrics, vol. 45 1, pp. 255–68, 1989

    work page 1989

  47. [47]

    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

  48. [48]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” in ICASSP, 2017

    work page 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.