Pith. sign in

REVIEW 2 major objections 4 minor 27 references

Layer-wise decision fusion in a frozen XLS-R model detects fake audio with 6.90% equal error rate on the In-the-Wild dataset, a cross-domain benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:00 UTC pith:7RM6TWQC

load-bearing objection Solid internal comparison and useful layer analysis, but the headline 6.90% EER overstates the cross-dataset claim because the baselines weren't run under the same protocol and the number is best-of-four on the test set. the 2 major comments →

arxiv 2607.20023 v1 pith:7RM6TWQC submitted 2026-07-22 cs.SD

Layer-Wise Decision Fusion for Fake Audio Detection Using XLS-R

Yixuan Xiao , Ngoc Thang Vu This is my paper
classification cs.SD
keywords fake audio detectiondeepfake speechlarge speech modelXLS-Rlayer-wise decision fusionone-class softmaxfeature collapsegeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that when a large pre-trained speech model like XLS-R is used for fake audio detection, fusing decisions made independently at each layer works better than fusing features from all layers into one representation before classification. The proposed layer-wise decision fusion (LW BN) achieves a cross-dataset equal error rate of 6.90% on the challenging In-the-Wild dataset, outperforming feature-fusion baselines and single-layer approaches. The paper also shows that this design makes the model more transparent, enabling analyses of how different layers contribute to detection, how silence in training data creates a shortcut, and what role discrete pseudo-phoneme tokens play.

Core claim

Attaching a one-class softmax classifier to each of the 25 XLS-R layer representations and then summing (or weighted-summing) the per-layer scores—rather than pooling all layer features into one vector before classification—prevents feature collapse and improves cross-dataset generalization. With a shared bottleneck projection and no reconstruction loss (LW BN), the method reaches an EER of 6.90% on In-the-Wild, compared to 10.97% for the feature-fusion variant under the same training protocol. Layer-weight analyses show that feature fusion tends to rely on a few inconsistently selected layers across runs, while decision fusion activates more layers and consistently emphasizes early-middle l

What carries the argument

Layer-wise decision fusion: each XLS-R layer's hidden representation is pooled by attention, projected through a shared bottleneck (optionally with an autoencoder-style reconstruction loss), and scored by its own one-class softmax classifier; the final score is the sum or learned weighted sum of the per-layer scores. This contrasts with feature fusion, which learns a softmax-normalized weight vector over layers to aggregate hidden states into a single representation before classification.

Load-bearing premise

The reported baselines (especially NN-ASP and NN-ACP) are directly comparable to the proposed models; the paper itself acknowledges that their feature-fusion reproduction (FF) uses fixed-length 4-second segments while those baselines use variable-length inputs, which could affect the performance gap.

What would settle it

Re-implement NN-ASP and NN-ACP under the same training protocol (fixed 4-second segments, same augmentation, same scheduler, same loss weighting) and compare EERs on In-the-Wild. If the gap between 6.90% and the re-implemented baselines vanishes or reverses, the claimed decision-fusion advantage is an artifact of protocol mismatch.

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

If this is right

  • If correct, late decision fusion is a simple drop-in replacement for feature fusion that improves cross-domain robustness without fine-tuning the large speech model.
  • The method's transparency allows practitioners to inspect which layers carry discriminative information for a given spoofing type, guiding which layers to keep or discard for specific deployment scenarios.
  • Training without leading/trailing silence is confirmed as essential for real-world generalization; including silence produces a silent-segment shortcut that inflates in-domain performance but hurts cross-dataset EER.
  • The released important token sets provide a starting point for linking discrete pseudo-phoneme tokens to detection-relevant acoustic cues.

Where Pith is reading between the lines

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

  • The consistent activation of early-middle and late layers across runs suggests that robust detection may require a mix of local acoustic detail and global prosodic or linguistic structure—an hypothesis the paper's layer-wise lens makes directly testable.
  • A testable extension would be applying the same decision-fusion design to other large speech models (WavLM, Whisper) to see whether the benefit generalizes beyond XLS-R.
  • The moderate correlation between important-token similarity and cross-domain EER hints that a token-aware training objective—encouraging the model to focus on tokens that transfer across datasets—could yield a further robustness gain.
  • Because the fusion weights are learned, the architecture could be adapted online to emphasize layers that fire correctly on a target domain's small labeled probe set.

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

2 major / 4 minor

Summary. The paper proposes layer-wise decision fusion for fake audio detection. Instead of aggregating hidden features from all XLS-R layers into one utterance-level representation before classification (feature fusion), the model attaches a one-class softmax classifier to each of the 25 layers and fuses per-layer scores either by summation or by learned weights. Two bottleneck variants (LW BN, LW BNR) and a weighted-score variant (LW BNW) are compared with a feature-fusion baseline (FF) and with published baselines. Models are trained on ASVspoof19 LA and evaluated on ASVspoof19 LA and In-the-Wild (ITW). The best reported result is LW BN with an ITW EER of 6.90%, which the abstract and conclusion describe as the best cross-dataset performance among strong baselines. The paper also presents analyses of layer-weight patterns, silence handling, and discrete-token similarity across datasets.

Significance. If the results hold, the paper makes a useful contribution: late decision fusion is a simple and plausible remedy for the risk that feature fusion collapses information from different depths of a large pre-trained speech model. The internal controlled comparison is a genuine strength: all four LW variants beat the paper's own FF baseline (10.97 -> 9.11–6.90) across five seeds, with standard deviations reported. The analysis of layer usage, silence, and discrete tokens is thoughtful and moves beyond a single EER number. The main limitation is that the headline 'best on ITW' rests on baseline numbers from [8] that were not re-run under the same protocol, and the best-performing variant is selected after seeing the test set.

major comments (2)
  1. [Section 4, Table 1] The claim that LW BN achieves the 'best cross-dataset performance on the In-the-Wild dataset (EER 6.90%) compared to other strong baselines' is not supported by a protocol-matched comparison. The authors state that FF is similar to NN-ASP and that the gap between FF (10.97) and NN-ASP (9.49) 'likely stems from' the fixed-length vs variable-length input. This is an explicit admission that the published NN-ASP/NN-ACP numbers come from a different training/inference protocol. Since the margin over NN-ASP is 2.59 EER, which is within the reported cross-run std of FF and could plausibly close under a matched protocol, the headline comparison is load-bearing and needs to be re-run under identical conditions (same fixed-length 4-second segments, same augmentation, same loss weighting, same early stopping) or explicitly re-labeled as non-comparable and removed from the abstract/conclusion.
  2. [Section 4, Table 1] The headline 6.90% is the best of four LW variants evaluated on the test set (LW 9.11, LW BN 6.90, LW BNR 7.40, LW BNW 7.52). Selecting the best variant after seeing ITW performance makes the reported number optimistic. This is not fatal for the qualitative claim that decision fusion beats feature fusion — all four variants outperform FF — but it does undermine the specific '6.90%' as a claim about a pre-specified method. The authors should either pre-register the selection rule, use a validation split for model selection, or explicitly state that 6.90% is the result of post-hoc selection over four variants.
minor comments (4)
  1. [Throughout] Typos and formatting: 'Utliizing' (Abstract), 'V oice' (Introduction), 'contribution 1 includes' should be 'contributions include', 'the followinglayer-wiseanalysis' is missing spaces, 'xlsr-r-300w' is inconsistently capitalized, and 'ASVpsoof19' is a typo. These should be cleaned up.
  2. [Section 5.2, Figure 4] Figure 4d has a missing title (blank after 'Fake'), and the four panels are not individually captioned. Please fix for readability.
  3. [Section 5.3, Figure 5] The x-axis similarity labels are hard to read; for example, the value '6.86' appears at the top of the first bar but is not clearly separated from the axis. The definition of 'important token set' (top 10% per trial per layer, then intersection across trials) would benefit from a one-sentence clarification in the figure caption.
  4. [Section 4, Table 1] The claim that LW variants 'consistently outperform' NN-ASP and NN-ACP is based on means only. Given that five seeds were run, a paired significance test or per-seed reporting would strengthen the statement, especially where standard deviations overlap.

Circularity Check

0 steps flagged

No circularity: the method is an empirical architecture comparison with no fitted-vs-predicted equation or self-citation chain.

full rationale

The paper contains no derivation chain in which an output is defined in terms of the claimed result. The proposed LW BN model is trained on ASVspoof19 LA with a one-class softmax loss and then evaluated on In-the-Wild; the ITW EER of 6.90% is a measured evaluation metric, not a value obtained by fitting to ITW labels or by renaming an input. The four LW variants (LW, LW BN, LW BNR, LW BNW) are all reported, and the qualitative conclusion that decision fusion outperforms feature fusion (FF 10.97% vs LW variants 6.90-9.11%) holds across the entire family, so the headline is not a single cherry-picked number that carries the argument alone. The authors' own note that NN-ASP/NN-ACP use variable-length input while the proposed models use fixed 4-second segments is a protocol-comparability limitation, not a circular step: it affects whether the external baselines are directly comparable, but it does not make the claimed result equivalent to its inputs. The analyses of layer weights, center similarities, and important discrete tokens are post-hoc explanatory observations derived from the trained model; they are not used as premises to define the architecture or to compute the reported EER. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The moderate token-similarity correlation is an exploratory finding, not a prediction forced by construction. Overall, no pattern of self-definitional, fitted-input-called-prediction, self-citation-load-bearing, or renaming circularity is present.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The method relies on standard one-class softmax classification, layer-wise feature diversity of XLS-R, and attention pooling; no exotic entities. The free-parameter list captures the hand-tuned choices (bottleneck dimension, H, 0.5 reconstruction weight, silence trimming, 1:9 loss weight, 10% token threshold) that the empirical result rests on. The main novel analytical construct—the 'important discrete token set'—is derived from existing XLS-R tokens and is not an invented entity.

free parameters (6)
  • Bottleneck dimension d = Not reported
    Dimension of the per-layer low-dimensional projection Wb∈R^{D×d}; controls classifier capacity and is a hand-tuned hyperparameter central to LW BN/LW BNR/LW BNW.
  • Number of attention heads H = 4
    Set to 4 after preliminary experiments (footnote 3) to avoid overfitting from 25 attention maps.
  • Reconstruction loss coefficient = 0.5
    Total loss = 0.5*MSELoss + average(classifier losses); no sensitivity analysis reported.
  • Silence trimming choice = without silence
    The paper reports both variants; the headline 6.90% uses 'without silence', selected because silent segments hurt ITW generalization.
  • Weighted loss ratio (fake:real) = 1:9
    Used to handle class imbalance; chosen by hand but standard in the field.
  • Important-token threshold (top 10%) = 10%
    Threshold defining the important token set in Analysis 3; correlation of -0.45 depends on this cut.
axioms (6)
  • domain assumption One-class softmax loss learns a compact real-audio center and generalizes to unseen spoofing attacks
    Adopted from [2] as the classifier for every layer; central to the method, restated in Section 2.1-2.2.
  • domain assumption Layers of XLS-R encode complementary acoustic, phonetic, and linguistic information
    Basis for attaching a classifier to every layer; supported by [9,10] but assumed for XLS-R in this task (Section 1).
  • domain assumption The attention pooling with a single affine map per layer is sufficient to aggregate frame-level evidence
    Section 2.2 design choice; no ablation over pooling methods.
  • standard math Autoencoder-style reconstruction loss keeps the shared projection informative
    Uses standard autoencoder idea [13]; treated as regularization without proof of benefit (Section 2.2).
  • domain assumption ASVspoof19 LA training data and In-the-Wild test data are a valid train/OOD-test pair for cross-dataset generalization
    Defines the evaluation protocol (Section 3); only one OOD set is used.
  • domain assumption XLS-R discrete tokens are meaningful pseudo-phonemes for the important-token analysis
    Section 5.3; the analysis highlights tokens with high attention scores and treats them as linguistic units, though [18] notes token-phoneme mapping is distributional.

pith-pipeline@v1.3.0-alltime-deepseek · 7970 in / 16477 out tokens · 147774 ms · 2026-08-01T11:00:44.457266+00:00 · methodology

0 comments
read the original abstract

Recent fake audio detection methods often leverage large speech models to achieve robust speech representations. These models are typically very deep, providing multiple layer-wise representations. However, current works often rely solely on single layer representation or feature fusion to extract one utterance-level representation for decision making. These methods risk underutilizing rich information from multiple layers and might induce feature collapse. We propose a novel layer-wise decision fusion method that applies fusion after per-layer decision making and achieves the best cross-dataset performance on In-the-Wild dataset (EER 6.90%) compared to other strong baselines. Our model design also makes the model more transparent, allowing us to conduct detailed analysis to reveal the underlying mechanism of decision making.

Figures

Figures reproduced from arXiv: 2607.20023 by Ngoc Thang Vu, Yixuan Xiao.

Figure 1
Figure 1. Figure 1: Model architecture. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Layer weight heatmap. Although layer weights are not available in LW BN, we an￾alyze the consecutive center similarity to study its behavior. In [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: , the value at layer i is the cosine similarity between the real center at layer i and that at layer i − 1. Results show that centers in layers 2–8 and 17–19 are highly similar across runs. The layer selection is similar to LW BNW. These similar cen￾ters can be viewed as the model automatically aggregates and augments important information (by summing their scores up). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1… view at source ↗
Figure 4
Figure 4. Figure 4: Cosine distance distribution across layers. have different noises and distortions, reducing generalization. When the model is trained without silence, more fake samples are misclassified as real, possibly because: 1) model trained on ASVspoof19 relies more on real samples’ characteristics to make a decision, 2) ITW’s low-quality real samples share less similarity with ASVpsoof19’s real samples, hence the m… view at source ↗
Figure 6
Figure 6. Figure 6: Example audio with highlighted important tokens. 6. Conclusion We propose a novel decision fusion method that has the best cross-dataset performance compared to baselines. Detailed analyses reveal that, unlike the feature fusion model which suf￾fers from feature collapse and has inconsistent layer reliance, our method includes features from more layers and shows con￾sistent reliance. Moreover, including si… view at source ↗
Figure 5
Figure 5. Figure 5: Token set similarity and cross-dataset performance. To further explain the inconsistency we observed, we per￾formed a qualitative analysis by focusing on the important frames and their corresponding tokens. Any frame with a token label that appears in the important token set is considered im￾portant and is highlighted in the waveform. We notice that even though some layers have more than 100 important toke… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

27 extracted references · 4 linked inside Pith

  1. [1]

    While modern FAD systems can achieve high in-domain performance, a main challenge lies in their generalization abil- ities [1]

    Introduction Fake audio detection (FAD) refers to the task of determining whether a given speech recording is genuine or synthetically generated (e.g., via Text-to-Speech (TTS) or V oice Conversion (VC)). While modern FAD systems can achieve high in-domain performance, a main challenge lies in their generalization abil- ities [1]. Specifically, the perfor...

  2. [2]

    Feature Fusion Methods We implement aFeature Fusion(FF) method based on [8] for comparative analysis

    Methods 2.1. Feature Fusion Methods We implement aFeature Fusion(FF) method based on [8] for comparative analysis. The method first extracts hidden repre- 1https://github.com/XIAOYixuan/tomatoDD/tree/ interspeech25-layer-wise arXiv:2607.20023v1 [cs.SD] 22 Jul 2026 sentationsh l ∈R T×D (D= 1024for XLS-R-300M 2) from a pretrained speech model for layerl. Th...

    Pith/arXiv arXiv 2026

  3. [3]

    The latter uses a learn- able weight vectorα∈R L, normalized via softmax

    by summing all scores or 2) by learning a weight for each layer and summing the weighted scores. The latter uses a learn- able weight vectorα∈R L, normalized via softmax. Mod- els using weighted scores are denoted with a ”W” suffix (e.g., LW BNW)

  4. [4]

    The ASVspoof19 LA contains clean, high-quality recordings with spoofed speech generated using TTS and VC methods

    Experiment Setting DatasetAll models are trained on the ASVspoof19 LA[14] train set and evaluated on both the ASVspoof19 LA evalua- tion set and the In-the-Wild (ITW) dataset. The ASVspoof19 LA contains clean, high-quality recordings with spoofed speech generated using TTS and VC methods. In contrast, ITW col- lects audio from real-world sources, featurin...

  5. [5]

    Results and Discussion Table 1:Performance comparison. Method ASV ITW Baseline Methods XLS-R+logres[7]: 300m (w silence) 1.00 21.30 1b (w silence) 1.30 18.70 2b (w silence) 0.60 7.20 wav2vec2+binary classifier [11] 2.98 26.65 NN-ASP (w silence) [8] 0.22 11.10 NN-ASP(w/o silence) [8] 5.56 9.49 NN-ACP (w silence) [8]0.1911.09 NN-ACP(w/o silence) [8] 8.09 10...

    1911

  6. [6]

    Conclusion We propose a novel decision fusion method that has the best cross-dataset performance compared to baselines. Detailed analyses reveal that, unlike the feature fusion model which suf- fers from feature collapse and has inconsistent layer reliance, our method includes features from more layers and shows con- sistent reliance. Moreover, including ...

  7. [7]

    However, as the model ca- pacity increases (e.g., from 300M to 2B parameter models used in [7]), performance improves noticeably

    and [11] use a single-layer representation and in general show poor cross-dataset performance, showing the limitation of using only one layer for detection. However, as the model ca- pacity increases (e.g., from 300M to 2B parameter models used in [7]), performance improves noticeably. Our proposed meth- ods, along with two baselinesNN-ASPandNN-ACP, which...

  8. [8]

    important

    Analysis In this section, we aim to understand the underlying reasons why LW variants outperform FF. Additionally, we conduct a layer-wise analysis to examine how silence affects performance. Does it only impact shallow layers? Are deeper layers, which are less relevant to simple signal characteristics, more resis- tant? Furthermore, since we can extract ...

  9. [9]

    activated

    with high average attention scores and 2) being “activated” repetitively across all trials. Discrete tokens can be assigned to each frame by XLS-R, and the frame’s attention score is gen- erated by our attention-based time pooling method during in- ference. Therefore, for a given dataset, token-score pairs can be collected per layer. Later, for each layer...

    arXiv

  10. [10]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P. Czempin, F. Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?” in 23rd Annual Conference of the International Speech Communi- cation Association, Interspeech 2022, Incheon, Korea, September 18-22, 2022, H. Ko and J. H. L. Hansen, Eds. ISCA, 2022, pp. 2783–2787

    2022

  11. [11]

    One-class learning towards synthetic voice spoofing detection,

    Y . Zhang, F. Jiang, and Z. Duan, “One-class learning towards synthetic voice spoofing detection,”IEEE Signal Process. Lett., vol. 28, pp. 937–941, 2021

    2021

  12. [12]

    One-class learning with adaptive centroid shift for audio deepfake detection,

    H. M. Kim, K. Jang, and H. Kim, “One-class learning with adaptive centroid shift for audio deepfake detection,”CoRR, vol. abs/2406.16716, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” inInternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, ser. Pro- ceedings of Machine Learning Research, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarl...

    2023

  14. [14]

    XLS-R: self-supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Con- neau, and M. Auli, “XLS-R: self-supervised cross-lingual speech representation learning at scale,” in23rd Annual Conference of the International Speech Communication Association, Interspeech 2022, Incheon, Korea, September 18-22, 20...

    2022

  15. [15]

    Wavlm: Large- scale self-supervised pre-training for full stack speech process- ing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “Wavlm: Large- scale self-supervised pre-training for full stack speech process- ing,”IEEE J. Sel. Top. Signal Process., vol. 16, no. 6, pp. 1505– 1518, 2022

    2022

  16. [16]

    Towards generalisable and calibrated audio deepfake detection with self- supervised representations,

    O. Pascu, A. Stan, D. Oneata, E. Oneata, and H. Cucu, “Towards generalisable and calibrated audio deepfake detection with self- supervised representations,” inInterspeech 2024, 2024, pp. 4828– 4832

    2024

  17. [17]

    Exploring self-supervised embeddings and syn- thetic data augmentation for robust audio deepfake detection,

    J. M. Mart ´ın-Do˜nas, A. ´Alvarez, E. Rosello, A. M. Gomez, and A. M. Peinado, “Exploring self-supervised embeddings and syn- thetic data augmentation for robust audio deepfake detection,” in Interspeech 2024, 2024, pp. 2085–2089

    2024

  18. [18]

    Layer-wise analysis of a self-supervised speech representation model,

    A. Pasad, J. Chou, and K. Livescu, “Layer-wise analysis of a self-supervised speech representation model,” inIEEE Automatic Speech Recognition and Understanding Workshop, ASRU 2021, Cartagena, Colombia, December 13-17, 2021. IEEE, 2021, pp. 914–921

    2021

  19. [19]

    Comparative layer-wise anal- ysis of self-supervised speech models,

    A. Pasad, B. Shi, and K. Livescu, “Comparative layer-wise anal- ysis of self-supervised speech models,” inIEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023. IEEE, 2023, pp. 1–5

    2023

  20. [20]

    Spoofed training data for speech spoofing countermeasure can be efficiently created using neu- ral vocoders,

    X. Wang and J. Yamagishi, “Spoofed training data for speech spoofing countermeasure can be efficiently created using neu- ral vocoders,” inIEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023. IEEE, 2023, pp. 1–5

    2023

  21. [21]

    Attentive merging of hidden embeddings from pre-trained speech model for anti- spoofing detection,

    Z. Pan, T. Liu, H. B. Sailor, and Q. Wang, “Attentive merging of hidden embeddings from pre-trained speech model for anti- spoofing detection,” inInterspeech 2024, 2024, pp. 2090–2094

    2024

  22. [22]

    Autoencoders,

    D. Bank, N. Koenigstein, and R. Giryes, “Autoencoders,” CoRR, vol. abs/2003.05991, 2020. [Online]. Available: https: //arxiv.org/abs/2003.05991

    Pith/arXiv arXiv 2003

  23. [23]

    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. Kinnunen, K. A. Lee, L. Juvela, P. Alku, Y . Peng, H. Hwang, Y . Tsao, H. 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. 64, p. 101114, 2020

    2019

  24. [24]

    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 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6369–6373

    2021

  25. [25]

    Rawboost: A raw data boosting and augmentation method ap- plied to automatic speaker verification anti-spoofing,

    H. Tak, M. R. Kamble, J. Patino, M. Todisco, and N. W. D. Evans, “Rawboost: A raw data boosting and augmentation method ap- plied to automatic speaker verification anti-spoofing,” inIEEE International Conference on Acoustics, Speech and Signal Pro- cessing, ICASSP 2022, Virtual and Singapore, 23-27 May 2022. IEEE, 2022, pp. 6382–6386

    2022

  26. [26]

    Speech is silver, silence is golden: What do asvspoof-trained models really learn?

    N. M. M ¨uller, F. Dieckmann, P. Czempin, R. Canals, and K. B ¨ottinger, “Speech is silver, silence is golden: What do asvspoof-trained models really learn?”CoRR, vol. abs/2106.12914, 2021. [Online]. Available: https://arxiv.org/abs/ 2106.12914

    Pith/arXiv arXiv 2021

  27. [27]

    An information-theoretic analysis of self-supervised discrete repre- sentations of speech,

    B. M. Abdullah, M. M. Shaik, B. M ¨obius, and D. Klakow, “An information-theoretic analysis of self-supervised discrete repre- sentations of speech,” in24th Annual Conference of the Inter- national Speech Communication Association, Interspeech 2023, Dublin, Ireland, August 20-24, 2023, N. Harte, J. Carson- Berndsen, and G. Jones, Eds. ISCA, 2023, pp. 2883–2887

    2023