Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

This paper claims that audio deepfake detectors are far less robust than average error-rate scores suggest, and that a new evaluation design—bona fide cross-testing—exposes worst-case equal error rates of 73–99% that standard benchmarks mis

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 ↗

A new evaluation protocol exhaustively pairs 164 speech synthesizers with nine bona fide speech types and reports max-pooled EERs, revealing larger failures than pooled averages show.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Bona fide cross-testing is a genuinely useful evaluation idea, but the paper's own numbers are unreliable until the metric equation, the table units, and the score-polarity question are fixed. the 4 major comments →

arxiv 2509.09204 v1 pith:6SIGV3V5 submitted 2025-09-11 cs.SD cs.AIcs.CL

Bona fide Cross Testing Reveals Weak Spot in Audio Deepfake Detection Systems

Chin Yuen Kwok , Jia Qi Yip , Zhen Qiu , Chi Hung Chi , Kwok Yan Lam This is my paper
classification cs.SD cs.AIcs.CL
keywords audio deepfake detectionbona fide cross-testingequal error rateevaluation methodologyspoof detectionmaximum poolingspeech anti-spoofinggeneralization
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

The authors propose a new way to evaluate audio deepfake detectors. Instead of melting many synthesizers into one test set and reporting a single error rate, they pair every synthesizer with nine different kinds of genuine speech and compute an equal error rate for each pair. Across 164 synthesizers, the maximum pooled error rate—the worst synthesizer for each genuine speech type—reaches 0.73 to 0.99 for three recent self-supervised detectors, while average error rates stay near 0.10. The paper argues that these worst-case numbers are the realistic ones because attackers will pick the easiest-to-miss synthesizer, and that genuine speech variety, not just synthesizer variety, drives the failures. The authors release their benchmark data and code so the field can use the same protocol.

Core claim

The central discovery is that the vulnerability of current audio deepfake detectors lies partly in the bona fide side of the test, not only in the spoof side. When the same set of 164 synthesizers is crossed with nine bona fide speech types—clean read speech, meetings, noisy accented speech, news, emotion, social media, and celebrity interviews—the equal error rates vary sharply by bona fide type. Two interview corpora (labeled b6 and b9) consistently produce the worst results, which the authors attribute to fast-paced celebrity speech recorded in noisy public areas. Averaging over synthesizers masks this: average EERs cluster near 10%, while maximum-pooled EERs reach 73–99%. The paper concl

What carries the argument

The key machinery is bona fide cross-testing combined with maximum pooling. Bona fide cross-testing takes K diverse genuine-speech datasets and M synthesizer datasets, forms M×K test sets, and computes a separate EER for each pair using the threshold that balances false positives and false negatives. Maximum pooling then summarizes the M EERs for each genuine speech type by taking the maximum, yielding mEER_k. This design counters two problems: underrepresented subsets lose influence in a combined dataset, and a single genuine-speech type gives no view of real-world diversity. The reported mEER_k values are the paper's main evidence for hidden vulnerabilities.

Load-bearing premise

The framework assumes the nine selected bona fide corpora, and the 600-sample cap per synthesizer–bona fide pair, capture real-world speech diversity and are directly comparable; systematic score offsets from differing channels, codecs, or speakers could make the b6/b9 weakness an artifact of dataset mismatch rather than genuine vulnerability.

What would settle it

Look for a dataset-identity artifact: compute the same mEER after per-corpus score normalization (e.g., z-scoring each bona fide subset's scores by its own mean and variance) or after matching duration and channel conditions; if the b6/b9 elevation shrinks toward the other types, the claimed genuine-speech vulnerability is an artifact. Alternatively, retrain a detector on interview-style bona fide data and check whether the max-pooled EER for b6/b9 drops substantially; it should if the paper's explanation is right.

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

If this is right

  • Average EER in the 10% range can coexist with worst-case per-synthesizer EERs above 0.9, so published averages understate risk.
  • Detectors trained on older benchmarks generalize poorly to newer synthesizers; several 2024 synthesizer families cause large errors.
  • Bona fide speech type is a first-order factor: celebrity interview speech (b6, b9) is markedly harder than clean read speech (b3, b5) for all three detectors.
  • Evaluating on combined multi-synthesizer sets can hide failures of underrepresented synthesizers because the global EER threshold is dominated by larger subsets.
  • Reporting per bona fide type is necessary for deployment decisions; for example, fake-news detection should use the mEER for news-domain bona fide audio.

Where Pith is reading between the lines

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

  • If bona fide cross-testing becomes standard, detector training should include diverse genuine speech, especially noisy conversational interview audio, rather than only clean read speech; this extension is implied but not developed in the paper.
  • A possible confound: EERs across corpora are compared without per-dataset score normalization; channel, codec, or speaker differences between b6/b9 and other corpora could contribute to the observed gaps. The paper does not test this.
  • The manuscript's reproducibility section contains a placeholder URL (https://empty.com) alongside the abstract's GitHub link; the stated benchmark release is not verifiable until that link resolves.
  • One testable extension: applying the same cross-test protocol to a detector trained with interview-style bona fide augmentation should shrink the b6/b9 gap, which would corroborate the authors' attribution to environmental difficulty.
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

4 major / 3 minor

Summary. The paper proposes 'bona fide cross-testing' for audio deepfake detection: instead of combining synthesizers into a single test set, each of 164 synthesizer subsets is paired with each of 9 bona fide speech corpora, and per-pair EERs are computed and then summarized by maximum and average across synthesizers. The authors argue that this reveals hidden vulnerabilities, especially on the celebrity-interview bona fide types b6 and b9, and that it is more robust and interpretable than traditional spoof cross-testing. They evaluate three SSL-based ADD models and claim to release a dataset, code, and score files.

Significance. The central idea addresses a real limitation of current ADD evaluation: pooling synthesizers in a single EER underweights underrepresented spoof subsets and ignores bona fide speech diversity. The construction of a benchmark with 164 synthesizers and 9 bona fide types is a useful contribution, and the proposal to report max-pooled EER is a reasonable way to surface worst-case performance. If the results were correct, the claim that current detectors are far less robust than average EERs suggest would be an important finding. However, the manuscript as written contains several load-bearing errors—an inconsistent definition of FPR in Eq. (8), an unresolved unit/polarity problem in Table 2, and a placeholder repository URL—that currently prevent the central evidence from being accepted.

major comments (4)
  1. [Section 4, Table 2] The table header says 'EER (%)' but the entries are decimals (e.g., 0.95), and the text interprets them as fractions ('fail to detect more than 30%', 'average EER approximately 10%'). Under the fraction reading, many max-EER values exceed 0.5 (e.g., 0.73, 0.98, 0.99). Equal error rate is bounded at 0.5 for a random or score-independent binary detector; values above 0.5 imply that the score ordering used in Eqs. (1)-(2) is anti-correlated with the class labels for those subsets. The paper never checks the polarity of the models' scores across the 164x9 pairs. If the polarity is reversed, the correct EER would be 1 - EER, and the 'hidden vulnerabilities' may largely disappear. If the entries are instead literal percentages, they are all below 1%, and the 'more than 30%' sentence is unsupported. Either way, the headline magnitude is not established.
  2. [Section 2.4, Eq. (8)] The definition of P^k_FP is inconsistent with Eq. (1). Eq. (8) uses the spoof-set symbol Lambda^k_P and the condition s_j >= tau, while Eq. (1) defines FPR over the bona fide set Lambda_N with s_j < tau. This is not a mere typo: Eqs. (9)-(10) use P^k_FP to compute EER^{k,m}, so if Eq. (8) is implemented literally, the framework computes a false-negative rate rather than a false-positive rate. The correct definition and the evaluation code need to be fixed and verified.
  3. [Section 3, data release] The abstract and contribution list promise code, datasets, and score files at https://github.com/cyaaronk/audio_deepfake_eval, but Section 3 says the material is 'available at https://empty.com'. This is a placeholder, not a usable repository. The reproducibility claim cannot be verified, and the reference to a nonexistent URL is a serious omission for a benchmark paper.
  4. [Section 4, b6/b9 claim] The claim that b6 and b9 are harder because they are 'celebrity interview speech... recorded in a noisy public area' is confounded by the dataset-level properties listed in Table 1: different sampling rates (44.1 kHz vs. 16/48 kHz), different durations, different codecs, and different recording setups. EERs computed on different corpora are compared directly without per-dataset score normalization or an analysis of score distribution shifts. It is therefore possible that the observed b6/b9 differences are artifacts of corpus-specific characteristics rather than intrinsic properties of the bona fide speech style. The paper should provide evidence that the effect persists after controlling for these factors.
minor comments (3)
  1. [General] Typos: 'Methodoglogy' in Section 2, 'nagative' in Section 2.2, 'accross' in Section 2.5, 'availabe' in Section 3, and 'AV-Deefake' in Table 1 should be corrected.
  2. [Section 2.5] The heading says 'Maximum pooling on spoof cross-testing results', but Eq. (11) pools over synthesizers for each bona fide type. The heading should reflect that this is max pooling over synthesizers within the bona fide cross-testing framework.
  3. [Table 1] The same source datasets (FakeAVCeleb and AV-Deepfake-1M) appear as both bona fide subsets (b6, b9) and spoof subsets (s3, s5). While this is a defensible choice, it should be stated explicitly and the implications for the 'cross-testing' terminology should be discussed.

Circularity Check

0 steps flagged

No significant circularity: bona fide cross-testing is an external benchmarking protocol, not a fitted prediction.

full rationale

The paper's contribution is an evaluation protocol, not a fitted predictive model. EER values in Table 2 are computed directly from model scores on external benchmark datasets (ASVspoof, LibriSpeech, VCTK, etc.) using the standard EER definition in Eqs. (1)-(4). No parameter is fitted to the reported outcomes and then re-predicted; maximum pooling in Eq. (11) is a summary statistic, not a derivation. Self-citations [1] and [28]-[30] appear only as examples of prior evaluation practice and as references for ASR/TTS corpora; they do not justify the central claim. The claim that b6/b9 are harder is an empirical observation from the computed grid, not an output forced by construction. Potential concerns about EER > 0.5 interpretation or dataset comparability are measurement/correctness issues, not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The paper does not introduce new physical entities, particles, or mediators. The 'bona fide cross-testing' framework is a protocol, not an entity. The axioms listed are the main modeling and measurement assumptions on which the evaluation conclusions rest.

free parameters (3)
  • samples_per_subset = 600
    The paper states it releases 600 audio samples per synthesizer and bona fide type. This cap is chosen without justification and affects EER stability.
  • num_bona_fide_types = 9
    K=9 bona fide datasets are selected ad hoc from available corpora; results and conclusions depend on this selection.
  • num_synthesizers = 164
    The benchmark includes 164 synthesizers from eight ADD datasets; this is not exhaustive and no sampling rationale is given.
axioms (5)
  • standard math EER is defined by a single threshold per test set that balances FPR and FNR (standard EER).
    Used in Eq (1)-(4) and throughout the evaluation.
  • domain assumption Attackers will use the synthesizer that is hardest to detect, so max pooling is the right summary.
    Justifies mEER in Section 2.5 and the focus on max values in Table 2.
  • domain assumption The nine bona fide datasets are representative of real-world speech diversity and are directly comparable.
    Section 4 concludes that interview speech (b6, b9) is harder by comparing EERs across datasets with different acoustics, speakers, and recording conditions.
  • domain assumption Pre-trained SSL model scores generalize across sampling rates after resampling.
    Section 3 states 'sampling rates are further standardized to match the requirements of each model', assuming this does not distort the scores.
  • domain assumption Combining subsets and computing a single EER is suboptimal because underrepresented subsets exert less influence on the threshold.
    This is the premise of the framework, motivated by Figure 1 and used to justify per-subset EER computation.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Bona fide Cross Testing Reveals Weak Spot in Audio Deepfake Detection Systems." pith.science (2026). https://pith.science/paper/6SIGV3V5

@misc{pith2026250909204,
  author       = {Pith},
  title        = {Pith review of: Bona fide Cross Testing Reveals Weak Spot in Audio Deepfake Detection Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6SIGV3V5}},
  note         = {Machine review of arXiv:2509.09204}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Audio deepfake detection (ADD) models are commonly evaluated using datasets that combine multiple synthesizers, with performance reported as a single Equal Error Rate (EER). However, this approach disproportionately weights synthesizers with more samples, underrepresenting others and reducing the overall reliability of EER. Additionally, most ADD datasets lack diversity in bona fide speech, often featuring a single environment and speech style (e.g., clean read speech), limiting their ability to simulate real-world conditions. To address these challenges, we propose bona fide cross-testing, a novel evaluation framework that incorporates diverse bona fide datasets and aggregates EERs for more balanced assessments. Our approach improves robustness and interpretability compared to traditional evaluation methods. We benchmark over 150 synthesizers across nine bona fide speech types and release a new dataset to facilitate further research at https://github.com/cyaaronk/audio_deepfake_eval.

Figures

Figures reproduced from arXiv: 2509.09204 by Chi Hung Chi, Chin Yuen Kwok, Jia Qi Yip, Kwok Yan Lam, Zhen Qiu.

Figure 1
Figure 1. Figure 1: Score distributions from three widely used data sub￾sets. The statistics of the subsets are shown in the top-left table. a) The EER for the green and orange subsets uses threshold −3.78. b) If the blue subset is used instead of the orange one, the threshold is 1.95. c) If the orange and blue subsets are com￾bined, the EER uses threshold −3.67. If only d) 10% or e) 1% of the orange subset is included in the… view at source ↗
Figure 2
Figure 2. Figure 2: Traditional evaluation results on 8 ADD testsets [2–6,26,27] by combining multiple synthesizers in a single test set. Limitations: (1) Existing ADD datasets typically contain a single bona fide speech type, limiting real-world diversity. (2) Some test sets, such as ASVspoof2021-DF, have an imbalanced distribution of audio samples across synthesizers. Underrepre￾sented subsets may exert less influence on th… 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 1 Pith paper

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

  1. Teffic-Audio: Tell Fact from Fiction

    cs.SD 2026-07 conditional novelty 4.0

    A simple Conformer deepfake detector trained with multi-source balanced sampling and diverse augmentation reaches 1.454% pooled EER on Speech-DF-Arena, first among public systems.

Reference graph

Works this paper leans on

41 extracted references · 12 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Introduction Audio deepfake detection (ADD) focuses on identifying syn- thetic or manipulated audio, commonly referred to as spoofed audio, which aims to replicate genuine recordings. These deep- fakes, created using advanced machine learning techniques, pose significant risks to systems relying on voice-based authen- tication, media forensics, and public...

    Pith/arXiv arXiv 2025

  2. [2]

    Despite this limita- tion, previous work [9,12,16] still merges synthesizer subsets into a single test set (e.g

    We show that combining multiple data subsets into a single test set is suboptimal because underrepresented subsets may have less influence on the EER threshold. Despite this limita- tion, previous work [9,12,16] still merges synthesizer subsets into a single test set (e.g. ASVspoof2021-DF [2]), although the subsets vary in size

  3. [3]

    No- tably, bona fide speech types contribute significantly to the challenges faced by ADD systems, whereas traditional ap- proaches primarily focus on the synthesizer type

    We reveal that bona fide cross-testing exposes vulnerabili- ties in ADD models that are often overlooked by conven- tional evaluation methods such as spoof cross-testing. No- tably, bona fide speech types contribute significantly to the challenges faced by ADD systems, whereas traditional ap- proaches primarily focus on the synthesizer type

  4. [4]

    Our approach demon- strates greater robustness and interpretability compared to traditional evaluation methods

    We propose a novel evaluation framework for ADD that in- tegrates bona fide cross-testing with spoof cross-testing and aggregates EER across spoof tests. Our approach demon- strates greater robustness and interpretability compared to traditional evaluation methods

  5. [5]

    We provide codes, datasets, and score files required to repli- cate benchmark results for over 150 synthesizers and 9 bona fide speech types. To the best of our knowledge, this is the first ADD evaluation framework that compares the perfor- mance of different bona fide speech types where both the acoustic environment and speech style varies

  6. [6]

    Methodoglogy This section begins by presenting the widely used spoof cross- testing evaluation framework. We then propose an extension to this framework, incorporating bona fide cross-testing, and pro- vide justifications for the modifications on the original evalua- tion approach based on the identified limitations. 2.1. Two types of errors: false positi...

  7. [7]

    In-The-Wild [25] is not included in the spoof test subsets as it uses unknown number of synthesizers

    Experiment Setup To improve the robustness of the evaluation, we collect audios from more than 150 synthesizers and 9 bona fide audio datasets as shown in Table 1. In-The-Wild [25] is not included in the spoof test subsets as it uses unknown number of synthesizers. Among the nine bona fide datasets, five of them are sourced from non-ADD corpora, originall...

    2006

  8. [8]

    As shown in Figure 2, all models perform well in datasets created in or before 2022

    Results and Discussions We begin by presenting the traditional spoof cross-testing eval- uation results on 8 ADD test sets. As shown in Figure 2, all models perform well in datasets created in or before 2022. Table 2:EER (%) results using our bona fide cross-testing evaluation framework. Given theM×KEERs obtained from bona fide and spoof cross-testing, we...

    2022

  9. [9]

    To support further research, we provide datasets and evaluation codes, aiming to drive the development of more robust and resilient audio deepfake detec- tion models

    Conclusion We introduce bona fide cross-testing, a novel evaluation frame- work that enhances robustness and interpretability by incorpo- rating diverse bona fide speech types, revealing vulnerabilities overlooked by traditional methods. To support further research, we provide datasets and evaluation codes, aiming to drive the development of more robust a...

  10. [10]

    Acknowledgements This research is supported by the National Research Founda- tion, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of Na- tional Research Foundation, Singap...

  11. [11]

    Robust audio deep- fake detection using ensemble confidence calibration,

    C. Y . Kwok, D.-T. Truong, and J. Q. Yip, “Robust audio deep- fake detection using ensemble confidence calibration,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  12. [12]

    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

  13. [13]

    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,”arXiv preprint arXiv:2401.09512, 2024

    Pith/arXiv arXiv 2024

  14. [14]

    Codecfake: Enhancing anti-spoofing models against deepfake audios from codec-based speech synthesis systems,

    H. Wu, Y . Tseng, and H.-y. Lee, “Codecfake: Enhancing anti-spoofing models against deepfake audios from codec-based speech synthesis systems,”arXiv preprint arXiv:2406.07237, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech,

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. W. D. Evans, M. Sahidullah, V . Vestman, T. H. Kinnunen, K. A. LEE, L. Juvela, P. Alku, Y .-H. Peng, H.-T. Hwang, Y . Tsao, H.-M. Wang, S. L. Maguer, M. Becker, and Z. Ling, “Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech,”Comput. Speech Lang., vol. 6...

    2019

  16. [16]

    Emofake: An initial dataset for emotion fake audio detection,

    Y . Zhao, J. Yi, J. Tao, C. Wang, and Y . Dong, “Emofake: An initial dataset for emotion fake audio detection,” inChina National Con- ference on Chinese Computational Linguistics. Springer, 2024, pp. 419–433

    2024

  17. [17]

    A method of estimating the equal error rate for automatic speaker verification,

    J.-M. Cheng and H.-C. Wang, “A method of estimating the equal error rate for automatic speaker verification,” in2004 In- ternational Symposium on Chinese Spoken Language Processing. IEEE, 2004, pp. 285–288

    2004

  18. [18]

    Learning a self- supervised domain-invariant feature representation for general- ized audio deepfake detection,

    Y . Xie, H. Cheng, Y . Wang, and L. Ye, “Learning a self- supervised domain-invariant feature representation for general- ized audio deepfake detection,” inProc. INTERSPEECH, vol. 2023, 2023, pp. 2808–2812

    2023

  19. [19]

    A conformer-based classifier for variable-length utterance pro- cessing in anti-spoofing,

    E. R. Casado, A. G. Alanıs, A. G. Garcıa, A. P. Herreroset al., “A conformer-based classifier for variable-length utterance pro- cessing in anti-spoofing,” inInt. Speech Conf.(INTERSPEECH), Dublin, Ireland, 2023

    2023

  20. [20]

    Multi-dataset co- training with sharpness-aware optimization for audio anti- spoofing,

    H.-j. Shim, J.-w. Jung, and T. Kinnunen, “Multi-dataset co- training with sharpness-aware optimization for audio anti- spoofing,”arXiv preprint arXiv:2305.19953, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Beyond silence: Bias analysis through loss and asymmetric approach in audio anti-spoofing,

    H.-j. Shim, M. Sahidullah, J.-w. Jung, S. Watanabe, and T. Kinnunen, “Beyond silence: Bias analysis through loss and asymmetric approach in audio anti-spoofing,”arXiv preprint arXiv:2406.17246, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    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

  23. [23]

    Data quality as predictor of voice anti-spoofing generalization,

    B. Chettri, R. G. Hautam ¨aki, M. Sahidullah, and T. Kinnunen, “Data quality as predictor of voice anti-spoofing generalization,” arXiv preprint arXiv:2103.14602, 2021

    Pith/arXiv arXiv 2021

  24. [24]

    Maxi- mum error modeling for fault-tolerant computation using maxi- mum a posteriori (map) hypothesis,

    K. Lingasubramanian, S. M. Alam, and S. Bhanja, “Maxi- mum error modeling for fault-tolerant computation using maxi- mum a posteriori (map) hypothesis,”Microelectronics Reliability, vol. 51, no. 2, pp. 485–501, 2011

    2011

  25. [25]

    A survey of emerging threats in cybersecurity,

    J. Jang-Jaccard and S. Nepal, “A survey of emerging threats in cybersecurity,”Journal of computer and system sciences, vol. 80, no. 5, pp. 973–993, 2014

    2014

  26. [26]

    Balance, mul- tiple augmentation, and re-synthesis: A triad training strategy for enhanced audio deepfake detection,

    T.-P. Doan, L. Nguyen-Vu, K. Hong, and S. Jung, “Balance, mul- tiple augmentation, and re-synthesis: A triad training strategy for enhanced audio deepfake detection,” inProc. Interspeech 2024, 2024, pp. 2105–2109

    2024

  27. [27]

    Iso/iec 30107: Information technology — biometric presentation attack detection,

    I. J. S. Biometrics, “Iso/iec 30107: Information technology — biometric presentation attack detection,” 2016

    2016

  28. [28]

    Range-based equal error rate for spoof localization,

    L. Zhang, X. Wang, E. Cooper, N. Evans, and J. Yamagishi, “Range-based equal error rate for spoof localization,” 2023. [Online]. Available: https://arxiv.org/abs/2305.17739

    Pith/arXiv arXiv 2023

  29. [29]

    ASVspoof 2015: the First Auto- matic Speaker Verification Spoofing and Countermeasures Chal- lenge,

    Z. Wu, T. Kinnunen, N. Evans, J. Yamagishi, C. Hanilc ¸i, M. Sahidullah, and A. Sizov, “ASVspoof 2015: the First Auto- matic Speaker Verification Spoofing and Countermeasures Chal- lenge,” inProc. Interspeech, 2015, pp. 2037–2041

    2015

  30. [30]

    A practical guide to logical access voice presentation attack detection,

    X. Wang and J. Yamagishi, “A practical guide to logical access voice presentation attack detection,” inFrontiers in Fake Media Generation and Detection. Springer, 2022, pp. 169–214

    2022

  31. [31]

    The ami meeting corpus: A pre-announcement,

    J. Carletta, S. Ashby, S. Bourban, M. Flynn, M. Guillemot, T. Hain, J. Kadlec, V . Karaiskos, W. Kraaij, M. Kronenthalet al., “The ami meeting corpus: A pre-announcement,” inInterna- tional workshop on machine learning for multimodal interaction. Springer, 2005, pp. 28–39

    2005

  32. [32]

    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,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  33. [33]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92),” 2019. [Online]. Available: https: //api.semanticscholar.org/CorpusID:213060286

    2019

  34. [34]

    Fakeavceleb: A novel audio-video multimodal deepfake dataset,

    H. Khalid, S. Tariq, and S. S. Woo, “Fakeavceleb: A novel audio-video multimodal deepfake dataset,”ArXiv, vol. abs/2108.05080, 2021. [Online]. Available: https: //api.semanticscholar.org/CorpusID:236976127

    Pith/arXiv arXiv 2021

  35. [35]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?” ArXiv, vol. abs/2203.16263, 2022. [Online]. Available: https: //api.semanticscholar.org/CorpusID:247793039

    arXiv 2022

  36. [36]

    Av-deepfake1m: A large-scale llm-driven audio- visual deepfake dataset,

    Z. Cai, S. Ghosh, A. P. Adatia, M. Hayat, A. Dhall, and K. Stefanov, “Av-deepfake1m: A large-scale llm-driven audio- visual deepfake dataset,” inACM Multimedia, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:265456085

    2023

  37. [37]

    Llamapartialspoof: An llm-driven fake speech dataset simulating disinformation generation,

    H.-T. Luong, H. Li, L. Zhang, K. A. Lee, and C. E. Siong, “Llamapartialspoof: An llm-driven fake speech dataset simulating disinformation generation,”ArXiv, vol. abs/2409.14743, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID: 272827910

    Pith/arXiv arXiv 2024

  38. [38]

    Continual learning op- timizations for auto-regressive decoder of multilingual asr sys- tems,

    C. Y . Kwok, J. Q. Yip, and E. S. Chng, “Continual learning op- timizations for auto-regressive decoder of multilingual asr sys- tems,”arXiv preprint arXiv:2407.03645, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Continual learning with embedding layer surgery and task- wise beam search using whisper,

    ——, “Continual learning with embedding layer surgery and task- wise beam search using whisper,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 140–146

    2024

  40. [40]

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

    C. Y . Kwok, H. Y . Li, and E. S. Chng, “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

  41. [41]

    Unsupervised cross-lingual representation learning for speech recognition,

    A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, “Unsupervised cross-lingual representation learning for speech recognition,” 2020. [Online]. Available: https: //arxiv.org/abs/2006.13979

    Pith/arXiv arXiv 2020

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.