Pith. sign in

REVIEW 4 major objections 6 minor 43 references

Two Views, One Truth: Spectral and Self-Supervised Features Fusion for Robust Speech Deepfake Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fusing two audio views cuts deepfake detection error by 38% on average, and reaches a new low on the In-The-Wild benchmark.

desk verdict Useful controlled ablation of fusion mechanisms, but headline gain rests on an unverified baseline from prior work. read the letter →

arxiv 2507.20417 v1 pith:GFVQQRUI submitted 2025-07-27 cs.SD cs.CReess.AS

Yassine El Kheir , Arnab Das , Enes Erdem Erdogan , Fabian Ritter-Guttierez , Tim Polzehl , Sebastian Möller This is my paper
classification cs.SDcs.CReess.AS
keywords audiodeepfakedetectionself-supervisedlearningspectralfeaturesfusioncross-attentionCQCCequalerrorrategeneralizationanti-spoofing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Audio deepfake detectors that rely only on self-supervised waveform embeddings or only on handcrafted spectral features each lose information. This paper argues that fusing the two views helps: it compares four fusion mechanisms and reports that every one beats the SSL-only baseline, with a cross-attention mechanism that lets Wav2Vec2.0 XLSR-53 embeddings selectively read Constant-Q cepstral coefficients (CQCC) reaching a 6.80% average equal error rate across four benchmarks. That is a 38% relative reduction from the 10.87% SSL-only baseline, and it includes an In-The-Wild (ITW) EER of 6.03% that the paper positions as a new best result among the systems it compares against. The paper's conclusion is that complementary spectral cues and rich contextual SSL representations jointly produce more domain-agnostic representations for detecting synthetic speech.

What carries the argument

The load-bearing mechanism is the cross-attention fusion (SSL $\rightarrow$ SF): aligned SSL and spectral feature matrices are projected to a common dimension, the SSL stream supplies queries $Q_{\mathrm{SSL}}$, the spectral stream supplies keys $K_{\mathrm{SF}}$ and values $V_{\mathrm{SF}}$, and the output $\mathrm{Softmax}(Q_{\mathrm{SSL}}K_{\mathrm{SF}}^{\top}/\sqrt{D})V_{\mathrm{SF}}$ is added to the SSL features through a residual connection. This lets the waveform-derived representation decide, per time step, which fine-grained cepstral cues are worth inheriting. The fused representation then feeds the AASIST graph-attention classifier, a named architecture that combines spectral and temporal information. The paper's interpretability analysis uses the learnable gating mechanism, a per-frame softmax weighting of the two streams, to quantify how much each view contributes.

What would settle it

Train the best configuration (Wav2Vec2.0 XLSR-53 plus CQCC with SSL-to-SF cross-attention) and the SSL-only baseline on the LA19 training split, then evaluate both on the official protocols of LA19, DF21, ITW, and ASV5. If the fused model's average EER does not stay clearly below the baseline's 10.87% or its ITW EER does not come in near 6.03%, the central claim fails. Re-running the cited comparison systems under this identical protocol would settle whether 6.03% is genuinely the new best ITW result.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a single lightweight cross-attention fusion module, in which self-supervised features are used as queries and handcrafted spectral features as keys and values, materially improves deepfake speech detection across domains. With CQCC as the spectral stream, the fused model achieves the best average EER of 6.80% over ASVspoof 2019 LA, ASVspoof 2021 DF, In-The-Wild, and ASVspoof 5, and the best EERs of 2.71% on DF21 and 6.03% on ITW among its four fusion variants. The paper also interprets the learnable gating variant to show that the network persistently assigns about 15% to 22% of its per-frame weight to the spectral stream and 78% to 85% to the SSL stream, with nearly identical distributions on two different datasets. The authors read this as evidence that the spectral branch contributes a stable, dataset-agnostic complement to the dominant SSL representation, and that the fusion is picking up a general-purpose policy rather than dataset-specific artifacts.

Load-bearing premise

The 'new best on ITW' claim assumes that the state-of-the-art numbers in the comparison table were obtained under the same training data, preprocessing, and evaluation protocols as this paper's results, an assumption the paper does not independently verify.

Editorial extensions

If this is right

  • All four fusion strategies—concatenation, cross-attention, mutual cross-attention, and learnable gating—beat the SSL-only baseline on every benchmark, with relative EER gains up to 48.77% on DF21, 57.00% on ITW, and 24.28% on ASV5.
  • Cross-attention with CQCC yields the best overall result, with average EER of 6.80% and per-dataset EERs of 0.40% on LA19, 2.71% on DF21, 6.03% on ITW, and 18.08% on ASV5.
  • Across all fusion schemes CQCC outperforms MFCC and LFCC on average, suggesting constant-Q cepstral descriptors carry the most complementary spectral detail for this task.
  • The gating analysis shows a stable allocation of roughly 15% to 22% of the fused representation to spectral features across datasets, so the fusion policy generalizes rather than memorizing dataset-specific artifacts.
  • The best model reaches competitive EERs without ensembles, multi-branch architectures, or dataset-specific augmentation, which is relevant for deployments that need one simple model across domains.

Reading between the lines

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

  • The same cross-attention routing should transfer to other SSL encoders such as WavLM or HuBERT and to other handcrafted descriptors such as formant trajectories or prosody, since the mechanism only requires aligned time and feature dimensions; this is an editorial extrapolation, not a claim tested in the paper.
  • Because the gating analysis shows a stable roughly 20% spectral contribution, a cheap testable extension would freeze or drop the spectral branch at inference time and measure how much of the measured gain survives.
  • The especially large gains on DF21 and ASV5, which contain codec-compressed and adversarially crafted audio, suggest the spectral branch is doing extra work under channel distortion; a direct test would be to compare fusion versus SSL-only after adding codecs or noise at test time.
  • A deployment-oriented extension would report tandem decision cost and calibration curves alongside EER, since EER equalizes false alarms and misses but real systems do not weight them equally.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes fusing self-supervised Wav2Vec2.0 XLSR-53 representations with handcrafted spectral features (MFCC, LFCC, CQCC) for speech deepfake detection. Four fusion strategies are compared—concatenation, cross-attention (SSL-to-SF), mutual cross-attention, and learnable gating—with AASIST as the backend classifier. Models are trained on ASVspoof 2019 LA train and evaluated on LA19, DF21, In-The-Wild, and ASVspoof 5. The central claim is that cross-attention fusion of CQCC with SSL features achieves the lowest average EER of 6.80%, a 38% relative reduction over the SSL-only baseline (10.87%), and a new best In-The-Wild result of 6.03%.

Significance. If the results are reproducible under a controlled protocol, the paper would offer a simple and interpretable fusion mechanism that improves cross-domain generalization in an important security application. The work is systematic in comparing three spectral descriptors and four fusion strategies on four public benchmarks, and the learnable-gating analysis provides insight into the relative contribution of each feature stream. However, the central quantitative claims currently depend on unverified baseline provenance and unspecified evaluation protocols. The paper does not report standard deviations despite three-seed runs, and no code or pretrained models are mentioned, which further limits the immediate value of the empirical claims. The significance is therefore conditional on resolving the baseline and protocol issues.

major comments (4)
  1. [Section 4.1, Table 1] The 'Baseline' row (0.28/5.29/14.03/23.88, average 10.87) is described as 'Wav2Vec2.0 XLSR-53 + AASIST [32]', but the DF21 and ITW values (5.29 and 14.03) are identical to the 'Wav2Vec2.0 AASIST' row attributed to [35] in Table 2. The manuscript does not state whether this baseline was trained by the authors under the same preprocessing and training recipe as the fusion models (pre-emphasis 0.97, 4-second truncation/padding, 201-frame SSL features, 50 epochs), nor does it give the source of the LA19 and ASV5 baseline entries. Because the headline 38% relative reduction is computed against this baseline, the comparison is controlled only if the baseline is rerun under identical conditions; otherwise the improvement may be an artifact of differing training setups. Please clarify the provenance and, if the numbers are quoted, either rerun the baseline or explicitly mark it as external and re-derive the relative improvement.
  2. [Section 3.2, Table 1] The paper states that each experiment was repeated three times with different random seeds, but Table 1 reports only point estimates with no standard deviations, confidence intervals, or per-seed values. Differences such as 6.80% vs 7.29% vs 8.09% between fusion strategies may be within seed variance, so the ranking of concatenation, cross-attention, mutual cross-attention, and gating is not statistically supported. Please report mean plus/minus standard deviation for the main comparison, and state whether the averages in Table 1 are means over seeds or results of a single run.
  3. [Section 4.3, Table 2] The claim of a new best result on In-The-Wild (6.03%) is made by direct comparison with EERs from other papers, but the evaluation protocol is not specified. The manuscript does not state which ITW trial list is used, how the 4-second truncation interacts with utterances shorter than 4 seconds, whether any score normalization is applied, or whether the ITW evaluation set is the full corpus or a subset. Since cited systems may use different training sets, augmentations (e.g., Rawboost), and evaluation splits, the numbers in Table 2 may not be protocol-comparable. Please document the exact ITW protocol, and ideally evaluate competing methods under the same protocol, before claiming state-of-the-art performance.
  4. [Section 3.1] Evaluation partitions are not defined precisely for any of the four benchmarks. For ASV5, the paper does not say whether the progress or eval subset is used; for DF21 and ITW, the official evaluation partitions and trial files are not referenced. Without this information, Table 1 cannot be reproduced and the cross-dataset average is not interpretable. Please add the exact partitions, trial lists, and any utterance-level filtering applied.
minor comments (6)
  1. [Introduction, Contributions] In the first contribution bullet, 'CFCC' appears to be a typo for 'CQCC'; the rest of the paper consistently uses CQCC.
  2. [Section 2.2] The downsampling of spectral features from 402 frames to 201 frames is not described; please specify the method (e.g., average pooling or strided sampling) and whether alignment uses interpolation.
  3. [Section 4.1] The citation [32] does not match a 'Wav2Vec2.0 + AASIST' baseline; the matching entry in Table 2 is [35]. Please correct the reference.
  4. [Section 3.2, Table 1] The 'No Fusion' setting is unclear: please state explicitly whether handcrafted features alone are also fed to AASIST and at what temporal resolution, so that these rows are reproducible.
  5. [Conclusion and Abstract] The claim that the fused representations are 'domain-agnostic' is too strong given four benchmarks of the same general kind; consider softening to 'more robust across the evaluated domains'.
  6. [Section 4.2, Figure 3] Figure 3 shows averaged normalized gating weights, but without error bars it is unclear whether the near-identical distributions on DF21 and ITW are stable across the three seeds; please add variability information or state that one representative seed is shown.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the fusion results are empirical comparisons on held-out benchmarks, and the only self-citation is non-load-bearing.

full rationale

The paper's central result—cross-attention fusion of Wav2Vec2.0 XLSR-53 and CQCC achieving a 6.80% average EER versus a 10.87% SSL-only baseline—is obtained by training on LA19 and evaluating on LA19, DF21, ITW, and ASV5. There is no equation in Sections 2.2 and 2.3 whose output is defined in terms of the reported EERs; the fusion weights are optimized with cross-entropy on training data, not fitted to test labels. The SSL branch uses official pretrained weights, and the spectral features are standard cepstral coefficients, none of which are fitted to the evaluation sets. The only self-citation, reference [27], supports the choice of Wav2Vec2.0 and does not enter the fusion derivation. A possible concern is that the Baseline row in Table 1 duplicates reference [35]'s DF21 and ITW numbers, so the '38% relative reduction' may rest on externally quoted rather than rerun baselines; that is a protocol or comparability issue, not circularity, because the fusion EERs themselves are measured, not derived from the baseline. Similarly, the SOTA positioning in Table 2 depends on published numbers with possibly different protocols, which is a correctness risk but not circularity. The derivation chain is therefore self-contained and empirically grounded.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities. The free parameters are standard ML hyperparameters and learned weights; the main ad-hoc choice is the temporal alignment between streams, which is load-bearing but unvalidated.

free parameters (5)
  • Common projection dimension D = 128
    Chosen by hand in Section 2.2; both feature streams are projected to 128 dimensions before fusion. This capacity choice affects all fusion variants and is not derived from data.
  • Training input length = 4 seconds (64,600 samples)
    Chosen in Section 3.2; truncates or zero-pads all audio, which can change feature distributions on datasets with different utterance lengths such as ITW (average 4.3 seconds).
  • Pre-emphasis coefficient = 0.97
    Chosen in Section 3.2; applied to all audio and could interact with spectral artifact capture.
  • Learnable per-frame gating weights = Learned; observed ~15-22% spectral weight
    Central to the interpretability analysis in Section 4.2; weights are optimized on LA19 only but their stability across datasets is used as evidence of a general fusion policy.
  • Training schedule = 50 epochs, batch 32, LR 1e-6, weight decay 1e-4
    Chosen by hand in Section 3.2; no tuning curves or sensitivity analysis are reported.
assumptions (4)
  • domain assumption Handcrafted spectral features contain spoofing artifacts complementary to SSL embeddings.
    Stated as motivation in Sections 1 and 2.1; the paper does not independently measure complementarity, it infers it from fusion gains.
  • domain assumption Training on ASVspoof 2019 LA and testing on DF21, ITW, and ASV5 measures cross-domain generalization.
    Section 3.1 defines the protocol; this assumes the evaluation sets are not seen during training and that EER on them is the right robustness metric.
  • domain assumption Wav2Vec2.0 XLSR-53 and AASIST are used as fixed building blocks whose pretrained parameters do not skew the comparison.
    Sections 2.1 and 2.3 rely on public pretrained weights and the AASIST architecture without ablating their influence.
  • ad hoc to paper Linear projection and temporal alignment of 402-frame spectral features to 201 SSL frames preserves the artifact information.
    Section 2.2 aligns TSF to TSSL with a linear projection; retaining temporal resolution is assumed to be sufficient, but the paper does not test alternative alignment strategies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Two Views, One Truth: Spectral and Self-Supervised Features Fusion for Robust Speech Deepfake Detection." pith.science (2026). https://pith.science/paper/GFVQQRUI

@misc{pith2026250720417,
  author       = {Pith},
  title        = {Pith review of: Two Views, One Truth: Spectral and Self-Supervised Features Fusion for Robust Speech Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GFVQQRUI}},
  note         = {Machine review of arXiv:2507.20417}
}
read the original abstract

Recent advances in synthetic speech have made audio deepfakes increasingly realistic, posing significant security risks. Existing detection methods that rely on a single modality, either raw waveform embeddings or spectral based features, are vulnerable to non spoof disturbances and often overfit to known forgery algorithms, resulting in poor generalization to unseen attacks. To address these shortcomings, we investigate hybrid fusion frameworks that integrate self supervised learning (SSL) based representations with handcrafted spectral descriptors (MFCC , LFCC, CQCC). By aligning and combining complementary information across modalities, these fusion approaches capture subtle artifacts that single feature approaches typically overlook. We explore several fusion strategies, including simple concatenation, cross attention, mutual cross attention, and a learnable gating mechanism, to optimally blend SSL features with fine grained spectral cues. We evaluate our approach on four challenging public benchmarks and report generalization performance. All fusion variants consistently outperform an SSL only baseline, with the cross attention strategy achieving the best generalization with a 38% relative reduction in equal error rate (EER). These results confirm that joint modeling of waveform and spectral views produces robust, domain agnostic representations for audio deepfake detection.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 29 canonical work pages

  1. [35]

    Automatic speaker verification spoofing and deepfake 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 deepfake detection using wav2vec 2.0 and data augmentation,” arXiv preprint arXiv:2202.12233 , 2022

    arXiv 2022

  2. [32]

    Generalizable speech deepfake detection via meta-learned LoRA

    J. Laakkonen, I. Kukanov, and V . Hautam¨aki, “Generalizable speech deep- fake detection via meta-learned lora,” arXiv preprint arXiv:2502.10838 , 2025

    work page Pith review arXiv 2025

  3. [1]

    A comprehensive survey with critical analysis for deepfake speech detection,

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

    work page 2025

  4. [2]

    Boosting large language model for speech synthesis: An empirical study,

    H. Hao, L. Zhou, S. Liu, J. Li, S. Hu, R. Wang, and F. Wei, “Boosting large language model for speech synthesis: An empirical study,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    work page 2025

  5. [3]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,

    D. Zhang, S. Li, X. Zhang, J. Zhan, P. Wang, Y . Zhou, and X. Qiu, “Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities,” arXiv preprint arXiv:2305.11000 , 2023

    arXiv 2023

  6. [4]

    Audiopalm: A large language model that can speak and listen,

    P. K. Rubenstein, C. Asawaroengchai, D. D. Nguyen, A. Bapna, Z. Borsos, F. d. C. Quitry, P. Chen, D. E. Badawy, W. Han, E. Kharitonov et al., “Audiopalm: A large language model that can speak and listen,” arXiv preprint arXiv:2306.12925, 2023

    arXiv 2023

  7. [5]

    Deepfake audio is a political nightmare,

    M. Meaker, “Deepfake audio is a political nightmare,” 2023

    work page 2023

  8. [6]

    Beware the artificial impostor: A mcafee cybersecurity artificial intelligence report 2023,

    McAfee, “Beware the artificial impostor: A mcafee cybersecurity artificial intelligence report 2023,” May 2023

    work page 2023

Show all 43 references
  1. [7]

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

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

    2021

  2. [8]

    The biometric vox system for the asvspoof 2021 challenge,

    J. C ´aceres, R. Font, T. Grau, J. Molina, and B. V . SL, “The biometric vox system for the asvspoof 2021 challenge,” in Proc. ASVspoof2021 Workshop, 2021

    2021

  3. [9]

    Known-unknown data augmentation strategies for detection of logical access, physical access and speech deepfake attacks: Asvspoof 2021,

    R. K. Das, “Known-unknown data augmentation strategies for detection of logical access, physical access and speech deepfake attacks: Asvspoof 2021,” Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge , pp. 29–36, 2021

    2021

  4. [10]

    Stc antispoofing systems for the asvspoof2021 challenge,

    A. Tomilov and A. e. a. Svishchev, “Stc antispoofing systems for the asvspoof2021 challenge,” in Proc. ASVspoof 2021 Workshop , 2021, pp. 61–67

    2021

  5. [11]

    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 preprint arXiv:2203.16263, 2022

    2022

  6. [12]

    A subnetwork approach for spoofing aware speaker verification

    A. Alenin, N. Torgashov, A. Okhotnikov, R. Makarov, and I. Yakovlev, “A subnetwork approach for spoofing aware speaker verification.” in INTERSPEECH, 2022, pp. 2888–2892

    2022

  7. [13]

    Pindrop labs’ submission to the asvspoof 2021 challenge,

    T. Chen, E. Khoury, K. Phatak, and G. Sivaraman, “Pindrop labs’ submission to the asvspoof 2021 challenge,” Proc. 2021 edition of the automatic speaker verification and spoofing countermeasures challenge , pp. 89–93, 2021

    2021

  8. [14]

    Multi-perspective information fusion res2net with randomspecmix for fake speech detection,

    S. Dong, J. Xue, C. Fan, K. Zhu, Y . Chen, and Z. Lv, “Multi-perspective information fusion res2net with randomspecmix for fake speech detection,” arXiv preprint arXiv:2306.15389 , 2023

    2023 arXiv

  9. [15]

    A comparative study on recent neural spoofing countermeasures for synthetic speech detection,

    X. Wang and J. Yamagishi, “A comparative study on recent neural spoofing countermeasures for synthetic speech detection,” arXiv preprint arXiv:2103.11326, 2021

    2021 arXiv

  10. [16]

    Multi-scale permutation entropy for audio deepfake detection,

    C. Wang, J. He, J. Yi, J. Tao, C. Y . Zhang, and X. Zhang, “Multi-scale permutation entropy for audio deepfake detection,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1406–1410

    2024

  11. [17]

    Raw differentiable architecture search for speech deepfake and spoofing detection,

    W. Ge, J. Patino, M. Todisco, and N. Evans, “Raw differentiable architecture search for speech deepfake and spoofing detection,” arXiv preprint arXiv:2107.12212, 2021

    2021 arXiv

  12. [18]

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

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks,” in ICASSP 2022-2022 IEEE . IEEE, 2022, pp. 6367–6371

    2022

  13. [19]

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

    Y . Xie, H. Cheng, Y . Wang, and L. Ye, “Learning a self-supervised domain-invariant feature representation for generalized audio deepfake detection,” in Proc. Interspeech, vol. 2023, no. 2023, 2023, pp. 2808– 2812

    2023

  14. [20]

    Can large-scale vocoded spoofed data improve speech spoofing countermeasure with a self-supervised front end?

    X. Wang and J. Yamagishi, “Can large-scale vocoded spoofed data improve speech spoofing countermeasure with a self-supervised front end?” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 311–10 315

    2024

  15. [21]

    A robust audio deepfake detection system via multi-view feature,

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

    2024

  16. [22]

    Harder or different? understanding generalization of audio deepfake detection,

    N. M. M ¨uller, N. Evans, H. Tak, P. Sperl, and K. B ¨ottinger, “Harder or different? understanding generalization of audio deepfake detection,” arXiv preprint arXiv:2406.03512 , 2024

    2024 arXiv

  17. [23]

    Combining spectral and self-supervised features for low resource speech recognition and translation,

    D. Berrebbi, J. Shi, B. Yan, O. L ´opez-Francisco, J. D. Amith, and S. Watanabe, “Combining spectral and self-supervised features for low resource speech recognition and translation,” arXiv preprint arXiv:2204.02470, 2022

    2022 arXiv

  18. [24]

    Fine-tune pre-trained models with multi-level feature fusion for speaker verification,

    S. Peng, W. Guo, H. Wu, Z. Li, and J. Zhang, “Fine-tune pre-trained models with multi-level feature fusion for speaker verification,” in Proc. Interspeech, 2024, pp. 2110–2114

    2024

  19. [25]

    Wave-spectrogram cross-modal aggregation for audio deepfake detection,

    Z. Jin, L. Lang, and B. Leng, “Wave-spectrogram cross-modal aggregation for audio deepfake detection,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  20. [26]

    Audio deepfake detection: What has been achieved and what lies ahead,

    B. Zhang, H. Cui, V . Nguyen, and M. Whitty, “Audio deepfake detection: What has been achieved and what lies ahead,” Sensors (Basel, Switzerland), vol. 25, no. 7, p. 1989, 2025

    1989

  21. [27]

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

    Y . E. e. a. Kheir, “Comprehensive layer-wise analysis of ssl models for audio deepfake detection,” arXiv preprint arXiv:2502.03559 , 2025

    2025 arXiv

  22. [28]

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

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . Vestman, T. Kinnunen, K. A. Leeet al., “Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech,” Computer Speech & Language , vol. 64, p. 101114, 2020

    2019

  23. [29]

    Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, J. Patino, A. Nautsch, X. Liu, K. A. Lee, T. Kinnunen, N. Evans et al. , “Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,” in ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoofing Cou...

    2021

  24. [30]

    Asvspoof 5: Crowdsourced speech 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. Kinnunen, N. Evans, K. A. Lee, and J. Yamagishi, “Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,” 2024. [Online]. Available: https://ar...

    2024 arXiv

  25. [31]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”

  26. [33]

    Constant q cepstral coefficients: A spoofing countermeasure for automatic speaker verification,

    M. Todisco, H. Delgado, and N. Evans, “Constant q cepstral coefficients: A spoofing countermeasure for automatic speaker verification,” Computer Speech & Language , vol. 45, pp. 516–535, 2017

    2017

  27. [34]

    Anti-spoofing ensembling model: Dynamic weight allocation in ensemble models for improved voice biometrics security,

    E. Rosell ´o Casado, ´A. M. G ´omez Garc ´ıa, I. L ´opez Espejo, A. M. Peinado Herreros, J. M. Mart ´ın Do ˜nas et al., “Anti-spoofing ensembling model: Dynamic weight allocation in ensemble models for improved voice biometrics security,” 2024

    2024

  28. [36]

    Spoofed speech detection with a focus on speaker embedding,

    H. M. Tran, D. Guennec, P. Martin, A. Sini, D. Lolive, A. Delhay, and P.-F. Marteau, “Spoofed speech detection with a focus on speaker embedding,” in INTERSPEECH 2024, 2024

    2024

  29. [37]

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

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

    2024

  30. [38]

    Mixture of experts fusion for fake audio detection using frozen wav2vec 2.0,

    Z. Wang, R. Fu, Z. Wen, J. Tao, X. Wang, Y . Xie, X. Qi, S. Shi, Y . Lu, Y . Liuet al., “Mixture of experts fusion for fake audio detection using frozen wav2vec 2.0,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2...

    2025

  31. [39]

    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

    2024 arXiv

  32. [40]

    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

    2025

  33. [41]

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

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

    2022

  34. [42]

    Opensmile: the munich versatile and fast open-source audio feature extractor,

    F. Eyben, M. W¨ollmer, and B. Schuller, “Opensmile: the munich versatile and fast open-source audio feature extractor,” in Proceedings of the 18th ACM international conference on Multimedia , 2010, pp. 1459–1462

    2010

  35. [2017]

    Available: https://arxiv.org/abs/1412.6980

    [Online]. Available: https://arxiv.org/abs/1412.6980

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.