Pith. sign in

REVIEW 3 major objections 7 minor 42 references

Wav2DF-TSL: Two-stage Learning with Efficient Pre-training and Hierarchical Experts Fusion for Robust Audio Deepfake Detection

T0 review · 3 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A two-stage pipeline that continues self-supervised training on unlabeled spoofed speech with lightweight adapters, then fuses hidden layers through a gated mixture of experts, reduces average equal error rate from 3.378% to 2.439% across f

desk verdict A solid, internally consistent engineering recipe for SSL-based audio deepfake detection with promising cross-domain gains, but the headline numbers are weakened by test-set tuning and an unreleased pretraining corpus. read the letter →

arxiv 2509.04161 v1 pith:O567A5D4 submitted 2025-09-04 cs.SD

Yunqi Hao , Yihao Chen , Minqiang Xu , Jianbo Zhan , Liang He , Lei Fang , Sian Fang , Lin Liu This is my paper
classification cs.SD
keywords audiodeepfakedetectionself-supervisedlearningparameter-efficientfine-tuningmixtureofexpertswav2vec2.0XLSRspoofedspeechpre-trainingcross-domaingeneralization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that the main weakness of self-supervised speech encoders in audio deepfake detection is not the classifier but the encoder's ignorance of what synthetic speech looks like, and that this can be repaired cheaply. It proposes a two-stage recipe: first continue training a wav2vec 2.0/XLSR encoder on 3,000 hours of unlabeled spoofed speech, updating only LoRA and convolutional adapter modules so the original speech knowledge is preserved; then fine-tune with a hierarchical adaptive mixture-of-experts that learns which transformer layers carry spoofing cues and fuses them through gated routing. On four benchmarks, the full system's average equal error rate drops from 3.378% for the baseline to 2.439%, a 27.8% relative improvement, with the largest gain on the out-of-domain In-the-wild set. The significance is that robustness to unseen attacks and codec distortions may be obtainable at the representation level rather than requiring a new classifier architecture.

What carries the argument

The central mechanism is a gating network whose expert-selection probabilities are modulated by a learned per-layer contribution vector. The vector is produced by statistically pooling all transformer hidden states, compressing them, and passing them through a sigmoid, so it encodes which of the 24 XLSR layers carry spoofing evidence. The weighted features are flattened and fed both to the gating network and to four small expert MLPs; the top-2 expert outputs are normalized and summed to produce the fused feature for the AASIST classifier. The pre-training stage uses LoRA low-rank matrices in attention plus a bottleneck 2D-convolutional adapter after attention and feed-forward blocks, so onl

What would settle it

Retrain the SSL-Hybrid stage while swapping the spoofed corpus for an equal amount of unlabeled real speech, or for data from only one generator family, keeping everything else fixed. If In-the-wild EER stays near the reported 7.762%, the claim that spoofed-speech pre-training drives generalization is false; if EER rises significantly, the claim is supported.

Watch

Extended reading notes

Core claim

The central claim is that exposing a self-supervised speech model to spoofed audio during its own pre-training objective, rather than only during supervised fine-tuning, makes the front-end features themselves more discriminant for deepfake detection. The paper implements this by freezing the XLSR weights and training only LoRA low-rank matrices and a bottleneck convolutional adapter inside each transformer block, with the standard wav2vec 2.0 contrastive loss on masked spoofed speech. It then adds a hierarchical adaptive mixture-of-experts layer that weights all 24 hidden layers with a learned contribution vector, uses those weighted features to route among four experts, and combines the to

Load-bearing premise

The load-bearing premise is that 3,000 hours of unlabeled spoofed speech from a specific set of generators teaches artifact patterns that transfer to attacks from generators and codec conditions the model has not seen; if the pre-training only memorizes generator-specific statistics, the cross-domain improvements would not hold outside this corpus.

Editorial extensions

If this is right

  • Continued self-supervised training on unlabeled spoofed speech works as a cheap domain-adaptation step: the hybrid LoRA+adapter version uses 29.68M trainable parameters against 317M for full fine-tuning and still beats full-parameter pre-training on average EER (2.826% vs 3.307%).
  • The HA-MoE module alone reduces average EER from 3.378% to 3.031%, so multi-layer feature selection contributes independently of spoofed pre-training.
  • Combining the two stages yields the best numbers: average EER 2.439%, with ASV-19LA at 0.103%, 21LA at 0.872%, 21DF at 1.954%, and In-the-wild at 6.827%.
  • Updating only adapters during fine-tuning reaches 2.718% average EER with 12% of the baseline parameters, implying the gains come from where updates are placed, not from parameter count alone.
  • On the In-the-wild set, the claimed 6.83% EER improves on prior published systems such as XLSR+SLS at 8.87% and one-class knowledge distillation at 7.68%, suggesting the method transfers to noisy, celebrity-speech settings.

Reading between the lines

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

  • If the mechanism is really artifact-level representation learning, the same two-stage recipe should transfer to other self-supervised encoders such as HuBERT or WavLM; the paper does not test this, but nothing in HA-MoE is XLSR-specific.
  • The construction of the 3,000-hour corpus mixes seven in-house generators with ASVSpoof5's 32 attack types, so the reported cross-domain gains may depend on that exact diversity; ablating each corpus component would reveal which portion is load-bearing.
  • A practical implication the paper leaves implicit is that, because only adapter and LoRA weights are trained in both stages, the method is well suited to rapidly updating a deployed detector when new synthesis families appear, without retraining the full encoder.
  • The HA-MoE gating vector could be inspected to see which transformer layers dominate for particular attack types; if layer preferences shift with codec and compression, the learned contribution vector is doing interpretable routing rather than generic pooling.
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

3 major / 7 minor

Summary. The paper proposes Wav2DF-TSL, a two-stage learning framework for audio deepfake detection (ADD). In the first stage, the XLSR self-supervised model is adapted to 3,000 hours of unlabeled spoofed speech using parameter-efficient modules (LoRA and a bottleneck convolutional adapter), with the claim that this learns artifact-sensitive representations while avoiding catastrophic forgetting. In the second stage, a hierarchical adaptive mixture-of-experts (HA-MoE) module fuses multi-layer hidden embeddings and feeds an AASIST classifier. The authors report EER improvements over a full-fine-tuning baseline on ASVspoof 2019 LA, 2021 LA, 2021 DF, and In-the-wild, including a 27.5% relative EER improvement on In-the-wild and 27.8% average relative improvement. The paper includes ablations showing the contributions of the SSL pre-training and the HA-MoE module, as well as comparisons with prior SOTA systems.

Significance. If the reported results are reliable, the proposed architecture is practically valuable: it achieves strong performance on four benchmarks, especially on the cross-domain In-the-wild dataset, while using parameter-efficient adaptation in the front end. The design is clearly specified and the ablation structure is logical. The use of a large unlabeled spoofed corpus and the HA-MoE fusion mechanism are interesting directions. However, the quantitative claims are currently weakened by two load-bearing issues: (i) hyperparameters are selected directly on the evaluation sets with no validation split, and (ii) the custom AudioFake corpus is unreleased and shares source corpora (VCTK, LJspeech) with the evaluation benchmarks, creating a potential source-overlap confound. These issues must be resolved before the SOTA claims can be accepted.

major comments (3)
  1. [§IV-A, Table I] The LoRA rank r and adapter size s are chosen by comparing EER on the four evaluation sets (ASV-19 LA, ASV-21 LA, ASV-21 DF, In-the-wild) and selecting r=8, s=64 as the best. No validation set is described anywhere in §III-B. This means the headline configuration, and therefore every subsequent result in Tables II–IV and the abstract, is selected using the test labels. This biases the reported improvements and invalidates the claim that the method 'significantly outperforms' the baseline. The authors should introduce a held-out validation split (e.g., a portion of the ASVSpoof19LA training set or a separate development set), tune r and s on that split, and then report evaluation-set results only for the final configuration. Sensitivity of the final results to the chosen hyperparameters should also be reported.
  2. [§III-A, AudioFake dataset] The AudioFake pre-training corpus is built from VCTK and LJspeech, and the ASVSpoof2019/2021 LA benchmarks also use VCTK-derived speech. This creates a direct source/recording-condition overlap between the pre-training data and the evaluation data. The paper does not examine this overlap or control for it. As a result, the reported gains on 19LA, 21LA, and 21DF could reflect speaker/channel memorization rather than transferable artifact representations. The In-the-wild result is less exposed to this confound, but the cross-domain claim is based on the whole system, and the source-overlap issue undermines the interpretation of the SSL pre-training stage. Please provide an analysis of speaker/source overlap, or an ablation in which the pre-training corpus is built from sources disjoint from the benchmark evaluation data, and report results separately for seen and unseen attack types.
  3. [§II-B, Eq. (4), and Table IV (A3 vs A1)] The first stage uses the standard wav2vec2.0 contrastive loss on unlabeled spoofed speech. The paper asserts that this 'learns artifacts' and improves generalization, but no evidence is provided that the resulting representations capture artifact-specific cues rather than generator-specific statistics or domain statistics. The ablation A3 vs A1 in Table IV compares the full fine-tuned system with and without the SSL pre-training stage, so any improvement could also come from the change in initialization and optimization trajectory during full fine-tuning (347.3M vs 317.8M parameters), not specifically from the learned embeddings. To make the claim load-bearing, please include controlled experiments: (i) linear-probe evaluation on frozen SSL features from the first stage, (ii) evaluation on held-out attack families not present in the AudioFake corpus, and (iii) a comparison pre-training o
minor comments (7)
  1. [Table IV] The header 'Full-Turning' is a typo for 'Full-Tuning'; also 'Full-Param' vs 'Full-Turning' is inconsistent between Table I and Table IV.
  2. [Eq. (4)] The contrastive loss formula is typeset incorrectly: the fraction line is missing, so the expression reads as 'log exp(...) P ...' instead of '-log [ exp(...) / sum ... ]'.
  3. [Eq. (5)] The notation 'ASPT' is not defined. Please define it explicitly as adaptive statistical pooling along the time dimension, and clarify the dimensions in Eq. (5) and Eq. (6).
  4. [§III-B, Implementation Details] No validation procedure is described for early stopping, learning-rate scheduling, or hyperparameter selection. Please state which data splits are used for validation and how early stopping is applied.
  5. [Figure 3] The t-SNE visualization is qualitative. Consider adding quantitative metrics (e.g., silhouette score or pairwise separation) to support the claim that Wav2DF-TSL produces a 'more compact and clearer decision boundary'.
  6. [Tables II and III] The comparisons to prior systems are not controlled: most prior systems do not have access to the 3,000-hour AudioFake corpus. Please clarify which comparisons are apples-to-apples and which reflect the additional pre-training data.
  7. [Abstract] The word 'significantly' is used without statistical significance testing. Given the small EER differences on some datasets (e.g., 0.872 vs 0.943 on 21LA), please report confidence intervals or multiple-run variability.

Circularity Check

1 steps flagged · score 4.0 of 10

Main headline gain is partly fitted by test-set configuration selection, but the core two-stage architecture has independent empirical content and no self-citation circularity.

  1. fitted input called prediction [Section IV-A, Table I; Section IV-C, Table IV]
    "Table I compares the performance of different self-supervised pretraining methods. ... Finally, by combining the LoRA and Adapter methods, we achieved the best results with an average EER of 2.826%. ... Experiment A4 demonstrates that combining the two-stage learning strategy, achieving an average EER of 2.439%. Compared to the baseline, it achieves a relative improvement of 27.8% on average, validating the compatibility and effectiveness of the proposed method."

    The final system's hyperparameters (LoRA r=8, Adapter s=64, hybrid) are selected as the best row of Table I, which is evaluated directly on the ASVSpoof 19LA/21LA/21DF and In-the-wild test sets. The same four test sets are then reused to report the headline '27.5% relative improvement' / 27.8% average improvement for the selected configuration in Table IV, and to claim state-of-the-art results in Tables II and III. Thus the reported gain is not an independent prediction for a fixed configuration; it is the best cell of a grid search performed on the evaluation target. The improvement is therefore partly an artifact of selecting on the test benchmarks rather than a pure consequence of the two-stage learning strategy. Some independent evidence remains (e.g., HA-MoE alone in A2), so the circu

full rationale

This paper is an empirical systems paper, not a formal derivation, so the classic 'Eq. X = Eq. Y by construction' circularity does not directly apply. I found no load-bearing self-citation chain: references such as XLSR-SLS, AASIST, and the ASVSpoof datasets are external prior work, and no 'uniqueness theorem' from the present authors is invoked. The proposed equations (LoRA update, adapter forward pass, gating network) define the architecture rather than assume the conclusion. The main circular element is methodological: Table I is an ablation/hyperparameter study run directly on the four evaluation benchmarks, the best configuration is picked from those test-set EERs, and the same test-set numbers are then presented as the final system's predictive performance and SOTA comparison. This is a test-set selection effect, not a definitional equivalence, and the central two-stage design still retains some independent support from ablations A2 and B2. The unverified generalization of the 3,000-hour pre-training corpus is a correctness/generalization risk, not circularity. Therefore a score of 4 is appropriate.

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

The paper's main free parameters are configuration choices, and several (r, s) are tuned directly on the test benchmarks, which is a circular element. The pre-training data is a custom, unreleased corpus, adding reliance on an unverifiable assumption.

free parameters (6)
  • LoRA rank r = 8
    Selected by comparing eval-set EER in Table I; no validation set described, so the choice is fitted to the test benchmarks.
  • Adapter bottleneck size s = 64
    Selected by comparing eval-set EER in Table I; s=128 degrades performance, showing test-set sensitivity.
  • Number of experts N = 4
    Fixed by hand without ablation in the paper.
  • Top-k = 2
    Fixed by hand without ablation in the paper.
  • Hierarchical bottleneck scale e = 12 (L/2)
    Design choice for the Vh computation in Eq. (6); not ablated.
  • Weighted CE class weights = 0.9/0.1
    Hand-set to address class imbalance; not ablated.
assumptions (4)
  • domain assumption XLSR-0.3B, pre-trained on 436k hours of real speech, is a suitable front-end for ADD after adaptation.
    The whole system builds on XLSR weights and assumes the SSL representations can be shifted to spoofed artifacts with adapters (Sec. II-B).
  • domain assumption The wav2vec 2.0 contrastive objective (Eq. 4) learns useful artifact representations when applied to unlabeled spoofed speech.
    The SSL pre-training stage continues training XLSR with the standard contrastive loss on the AudioFake corpus; if this loss does not capture spoofing artifacts, the first stage would not help.
  • ad hoc to paper The AudioFake corpus, generated with seven TTS/VC algorithms plus ASVSpoof5, is representative of spoofing artifacts in the evaluation benchmarks.
    The corpus is constructed by the authors and not released; the generalization claim depends on its coverage of artifact types (Sec. III-A).
  • domain assumption AASIST classifier is a reliable fixed classifier for the final prediction stage.
    The paper uses AASIST as-is and does not ablate it; performance is coupled to the classifier's behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wav2DF-TSL: Two-stage Learning with Efficient Pre-training and Hierarchical Experts Fusion for Robust Audio Deepfake Detection." pith.science (2026). https://pith.science/paper/O567A5D4

@misc{pith2026250904161,
  author       = {Pith},
  title        = {Pith review of: Wav2DF-TSL: Two-stage Learning with Efficient Pre-training and Hierarchical Experts Fusion for Robust Audio Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O567A5D4}},
  note         = {Machine review of arXiv:2509.04161}
}
read the original abstract

In recent years, self-supervised learning (SSL) models have made significant progress in audio deepfake detection (ADD) tasks. However, existing SSL models mainly rely on large-scale real speech for pre-training and lack the learning of spoofed samples, which leads to susceptibility to domain bias during the fine-tuning process of the ADD task. To this end, we propose a two-stage learning strategy (Wav2DF-TSL) based on pre-training and hierarchical expert fusion for robust audio deepfake detection. In the pre-training stage, we use adapters to efficiently learn artifacts from 3000 hours of unlabelled spoofed speech, improving the adaptability of front-end features while mitigating catastrophic forgetting. In the fine-tuning stage, we propose the hierarchical adaptive mixture of experts (HA-MoE) method to dynamically fuse multi-level spoofing cues through multi-expert collaboration with gated routing. Experimental results show that the proposed method significantly outperforms the baseline system on all four benchmark datasets, especially on the cross-domain In-the-wild dataset, achieving a 27.5% relative improvement in equal error rate (EER), outperforming the existing state-of-the-art systems. Index Terms: audio deepfake detection, self-supervised learning, parameter-efficient fine-tuning, mixture of experts

Figures

Figures reproduced from arXiv: 2509.04161 by the authors.

Figure 1
Figure 1. The framework of parameter-efficient self-supervised pretraining. (a) and (b) represent the in-block adapter network and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of the proposed ADD system. (a) represents the pipline of the fine-tuning phase of Wav2DF-TSL. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of t-SNE embeddings from the In-the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 39 canonical work pages

  1. [1]

    Asvspoof 2019: Future horizons in spoofed and fake audio detection,

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T. Kinnunen, and K. A. Lee, “Asvspoof 2019: Future horizons in spoofed and fake audio detection,” in Interspeech, 2019

    work page 2019

  2. [2]

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

    J. Yamagishi, X. Wang, M. Todisco, M. Sahidullah, and et.al, “Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,” in ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoof- ing Coutermeasures Challenge , 2021

    work page 2021

  3. [3]

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

    X. Wang, H. Delgado, H. Tak, J.-w. Jung, H.-j. Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunen, et al. , “Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,” arXiv preprint arXiv:2408.08739 , 2024

    arXiv 2024

  4. [4]

    Robust audio anti-spoofing with fusion-reconstruction learning on multi-order spectrograms,

    P. Wen, K. Hu, W. Yue, S. Zhang, W. Zhou, and Z. Wang, “Robust audio anti-spoofing with fusion-reconstruction learning on multi-order spectrograms,” in Interspeech, 2023

    work page 2023

  5. [5]

    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,” inInterspeech, pp. 4259–4263, 2021

    work page 2021

  6. [6]

    Channel-wise gated res2net: Towards robust detection of synthetic speech attacks,

    X. Li, X. Wu, H. Lu, X. Liu, and H. Meng, “Channel-wise gated res2net: Towards robust detection of synthetic speech attacks,” in Interspeech, pp. 4314–4318, 2021

    work page 2021

  7. [7]

    Fastaudio: A learnable audio front-end for spoof speech detection,

    Q. Fu, Z. Teng, J. White, M. G. Powell, and D. C. Schmidt, “Fastaudio: A learnable audio front-end for spoof speech detection,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 3693–3697, 2022

    work page 2022

  8. [8]

    The effect of silence and dual-band fusion in anti-spoofing system,

    Y . Zhang, W. Wang, and P. Zhang, “The effect of silence and dual-band fusion in anti-spoofing system,” in Interspeech, p. 4279–4283, 2021

    work page 2021

Show all 42 references
  1. [9]

    Towards end-to-end synthetic speech detection,

    G. Hua, A. Teoh, and H. Zhang, “Towards end-to-end synthetic speech detection,” IEEE Signal Processing Letters , vol. 28, pp. 1265–1269, 2021

    2021

  2. [10]

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

    J.-w. Jung, H.-S. Heo, H. Tak, and et al., “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 6367–6371, 2022

    2022

  3. [11]

    Discriminative frequency information learning for end-to-end speech anti-spoofing,

    B. Huang, S. Cui, J. Huang, and X. Kang, “Discriminative frequency information learning for end-to-end speech anti-spoofing,” IEEE Signal Processing Letters, vol. 30, pp. 185–189, 2023

    2023

  4. [12]

    Robust data2vec: Noise-robust speech representation learning for asr by com- bining regression and improved contrastive learning,

    Q. shi Zhu, L. Zhou, J. Zhang, S. Liu, Y . Hu, and L. Dai, “Robust data2vec: Noise-robust speech representation learning for asr by com- bining regression and improved contrastive learning,” pp. 1–5, 2023

    2023

  5. [13]

    Self-supervised learning with cluster-aware-dino for high-performance robust speaker verification,

    B. Han, Z. Chen, and Y . Qian, “Self-supervised learning with cluster-aware-dino for high-performance robust speaker verification,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, pp. 529–541, 2023

    2023

  6. [14]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Advances in Neural Information Processing Systems , pp. 12449–12460, 2020

    2020

  7. [15]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “Hubert: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language processing , vol. 29, pp. 3451–3460, 2021

    2021

  8. [16]

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

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, et al. , “Wavlm: Large-scale self-supervised pre- training for full stack speech processing,” IEEE Journal of Selected Topics in Signal Processing , pp. 1505–1518, 2022

    2022

  9. [17]

    Robust spoof speech detection based on multi-scale feature aggregation and dynamic convolution,

    H. Wu, J. Zhang, Z. Zhang, W. Zhao, B. Gu, and W. Guo, “Robust spoof speech detection based on multi-scale feature aggregation and dynamic convolution,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 10156–10160, 2024

    2024

  10. [18]

    Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmenta- tion,

    J. W. Jung, H.-S. Heo, H. Tak, et al. , “Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmenta- tion,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

    2022

  11. [19]

    Investigating self-supervised front ends for speech spoofing countermeasures,

    X. Wang and J. Yamagishi, “Investigating self-supervised front ends for speech spoofing countermeasures,” in The Speaker and Language Recognition Workshop (Odyssey), pp. 100–106, 2022

    2022

  12. [20]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, and et al., “Lora: Low-rank adaptation of large language models,” in International Conference on Machine Learning (ICML), 2022

    2022

  13. [21]

    Parameter-efficient transfer learning for nlp,

    A. C. Stickland and I. Murray, “Parameter-efficient transfer learning for nlp,” in International Conference on Machine Learning (ICML) , 2019

    2019

  14. [22]

    Audio deepfake detection with self- supervised xls-r and sls classifier,

    Q. Zhang, S. Wen, and T. Hu, “Audio deepfake detection with self- supervised xls-r and sls classifier,” in ACM Multimedia, 2024

    2024

  15. [23]

    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,” in Interspeech, pp. 2090–2094, 2024

  16. [24]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, and et al., “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” in International Conference on Learning Representations (ICLR) , 2017

    2017

  17. [25]

    Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,

    J. Kim, J. Kong, and J. Son, “Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,” in International Joint Conference on Artificial Intelligence , p. 5530–5540, 2021

    2021

  18. [26]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” in Advances in Neural Information Processing Systems , pp. 17022–17033, 2020

    2020

  19. [27]

    Fastdiff: A fast conditional diffusion model for high-quality speech synthesis,

    R. Huang, M. W. Y . Lam, J. Wang, and et al., “Fastdiff: A fast conditional diffusion model for high-quality speech synthesis,” in International Joint Conference on Artificial Intelligence , 2022

    2022

  20. [28]

    Freevc: Towards high-quality text-free voice conversion,

    W.-C. Huang, H.-Y . Lee, H.-Y . Liu, Y . Tsao, and H.-Y . Lee, “Freevc: Towards high-quality text-free voice conversion,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 1–5, 2023

    2023

  21. [29]

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

    X. Wang, J. Yamagishi, M. Todisco, and et al., “Asvspoof 2019: A large- scale public database of synthesized, converted and replayed speech,” Comput. Speech Lang. , vol. 64, p. 101114, 2019

    2019

  22. [30]

    Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,

    X. Liu, X. Wang, M. Sahidullah, and et al., “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 31, pp. 2507– 2522, 2023

    2021

  23. [31]

    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

  24. [32]

    t-dcf: a detection cost function for the tandem assessment of spoofing countermeasures and automatic speaker verification,

    T. Kinnunen, K. A. Lee, H. Delgado, and N. E. .et.al, “t-dcf: a detection cost function for the tandem assessment of spoofing countermeasures and automatic speaker verification,” in The Speaker and Language Recognition Workshop(Odyssey), 2018

    2018

  25. [33]

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

    H. Tak, M. Kamble, J. Patino, and et al., “Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti- spoofing,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 6382–6386, 2022

    2022

  26. [34]

    Improving short utterance anti-spoofing with aasist2,

    Y . Zhang, J. Lu, Z. Shang, W. Wang, and P. Zhang, “Improving short utterance anti-spoofing with aasist2,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 11636– 11640, 2024

    2024

  27. [35]

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

    E. Rosello, A. Gomez-Alanis, A. M. Gomez, and A. Peinado, “A conformer-based classifier for variable-length utterance processing in anti-spoofing,” in Interspeech, 2023

    2023

  28. [36]

    Audio deepfake detection with self-supervised wavlm and multi-fusion attentive classi- fier,

    Y . Guo, H. Huang, X. Chen, H. Zhao, and Y . Wang, “Audio deepfake detection with self-supervised wavlm and multi-fusion attentive classi- fier,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 12702–12706, 2024

    2024

  29. [37]

    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,” in Interspeech, 2024

    2024

  30. [38]

    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 C. E. Siong, “Temporal-channel modeling in multi-head self-attention for synthetic speech detection,” 2024

    2024

  31. [39]

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

    Y . Yang, H. Qin, H. Zhou, C. Wang, and et al., “A robust audio deepfake detection system via multi-view feature,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 13131– 13135, 2024

    2024

  32. [40]

    Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,

    X. Wang and J. Yamagishi, “Spoofed training data for speech spoofing countermeasure can be efficiently created using neural vocoders,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1–5, 2023

    2023

  33. [41]

    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?,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 10311–10315, 2023

    2023

  34. [42]

    One-class knowl- edge distillation for spoofing speech detection,

    J. Lu, Y . Zhang, W. Wang, Z. Shang, and P. Zhang, “One-class knowl- edge distillation for spoofing speech detection,” IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pp. 11251–11255, 2024

    2024

Pith tools

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