Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

MoLEx: Mixture of LoRA Experts in Speech Self-Supervised Models for Audio Deepfake Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Mixture of LoRA experts with a learned router detects audio deepfakes at 5.56% error on ASVSpoof 5 while cutting trainable parameters by 86.4%.

desk verdict MoLEx is a sensible parameter-efficient recipe with a thorough ablation, but the 5.56% EER headline is inflated by eval-set hyperparameter selection and the SOTA claim overreaches. read the letter →

arxiv 2509.09175 v1 pith:OL3HTKHK submitted 2025-09-11 cs.SD cs.MM

Zihan Pan , Sailor Hardik Bhupendra , Jinyang Wu This is my paper
classification cs.SDcs.MM
keywords audiodeepfakedetectionself-supervisedlearningWavLMLoRAmixtureofexpertsparameter-efficientfine-tuningASVSpoof5orthogonalityregularization
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

This paper tries to establish that a mixture of low-rank adapter experts, gated by a learned router, can adapt a frozen self-supervised speech model to detect audio deepfakes as accurately as full fine-tuning while using far fewer trainable parameters. On the ASVSpoof 5 benchmark the authors report an equal error rate of 5.56% on the evaluation set, which they compare favorably to other single-model results. The paper further argues that the router learns domain-aware specialization: it re-uses the same experts for attacks from the same synthesis family and shifts to different experts for novel spoofs. It also shows that new experts can be added for domain adaptation without retraining the whole model. If true, this would make state-of-the-art deepfake detection much cheaper and easier to keep current as new synthesis engines appear.

What carries the argument

The key machinery is the MoLEx module: in each transformer layer, N low-rank LoRA adapters (each a product of two small matrices) sit in parallel with the frozen feed-forward network, and a noisy gating network computes scores that select the top K adapters for each input. The weighted sum of the selected adapters' outputs is added to the frozen FFN output. An orthogonality regularization loss, computed as the Frobenius distance between the Gram matrix of each adapter and the identity, keeps the adapters from collapsing to a lower effective rank.

What would settle it

Hold out the ASVSpoof 5 evaluation set completely; choose all hyperparameters (rank, K/N, number of layers, number of experts) using only the development set, then measure the evaluation error rate once. If the result is substantially higher than 5.56%, the reported number is inflated by test-set selection.

Watch

Extended reading notes

Core claim

The central discovery is that replacing the frozen feed-forward weight matrix in each WavLM transformer layer with a sparse mixture of low-rank (LoRA) adapters, selected by a noisy top-K router, yields state-of-the-art audio deepfake detection without updating the backbone. The authors add an orthogonality regularizer that keeps the effective rank of each LoRA adapter high, preventing rank collapse and improving performance. They also show, through gating-score heatmaps, that the router's expert choices are consistent for spoofing attacks that share the same synthesis lineage and change when the attack family differs. Finally, they demonstrate that adapting to a new domain can be done by fre

Load-bearing premise

The 5.56% evaluation-set error rate is treated as a fair measure of generalization, even though the configuration that produced it was chosen by looking at that same evaluation set.

Editorial extensions

If this is right

  • MoLEx reduces trainable parameters by about 86% compared with full fine-tuning of WavLM-Large while matching or beating full-finetuning error rates on ASVSpoof 5.
  • The router's expert selection is interpretable and domain-aware: similar attack families reactivate the same experts, while novel attacks switch experts.
  • The orthogonality regularizer prevents rank collapse and improves error rate across settings, e.g., from 7.04% to 5.56% in the best configuration.
  • Domain adaptation can be achieved by training only new LoRA experts and routers, with a small amount of source-domain data reducing forgetting.
  • In cross-dataset evaluation, the same MoLEx configuration beats published single-system baselines on In-the-Wild, LibriSeVoc, DFADD, and Fake-or-Real when trained on the respective source corpus.

Reading between the lines

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

  • Because the 5.56% number came from a configuration selected after looking at the evaluation-set scores in Table III, the true generalization gain over other configurations is likely smaller than the paper implies; using the development set for model selection and then evaluating once would give an unbiased estimate.
  • The expert-utilization heatmaps suggest a practical side-channel: monitoring which experts the router activates could serve as an attack-family fingerprint or as an out-of-distribution alert when novel spoofs arrive.
  • The adaptation experiment hints at a deployment strategy where a base detector is frozen and only a handful of experts are added per emerging synthesis engine, which could be tested against a chronological stream of new TTS systems.
  • The orthogonality loss is a general trick for low-rank adapters and could help in other parameter-efficient transfer settings, though the paper does not test it beyond speech.
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 / 6 minor

Summary. The paper proposes MoLEx (Mixture of LoRA Experts), a parameter-efficient fine-tuning framework for speech SSL models, applied to audio deepfake detection. LoRA adapters are inserted in parallel with the FFN of transformer layers in a frozen WavLM-Large, a noisy-gating router selects the top-K experts per input, and an orthogonality regularization loss is introduced to preserve the effective rank of LoRA experts. The authors report a 5.56% EER on the ASVSpoof 5 evaluation set (Track 1) without data augmentation, claim state-of-the-art performance, and further present cross-dataset experiments and a domain adaptation scenario where only new experts and routers are trained.

Significance. If the results hold, the paper makes a useful contribution to parameter-efficient deepfake detection: the architecture is clearly described, the orthogonality regularization is a sensible mechanism to avoid rank collapse, and the code is released, which supports reproducibility. The expert-utilization analysis is a nice qualitative addition. However, the headline SOTA claim depends on a number that appears to have been selected on the evaluation set, and the same table used for the SOTA comparison internally contradicts the claim. These issues are load-bearing for the paper's central contribution, but they are correctable within the manuscript's scope.

major comments (3)
  1. [Table III and Section IV-B] The headline 5.56% EER is not a valid unbiased estimate of generalization because the hyperparameter configuration appears to have been selected on the evaluation set. Table III shows that configurations with lower development EER (0.37% for K/N=2/12, 0.55% for 12/12, 0.78% for 2/6 and 8/12) have eval EERs of 7.60%, 5.78%, 6.30%, and 5.69%, respectively, all worse than 5.56%. The configuration reported as best (K/N=4/12, r=32) has a dev EER of 1.25% but the lowest eval EER. Section III-B says 'we fix the optimal hyperparameter setting' after presenting Table III, but the selection criterion is not described. The pattern indicates that eval-set performance was used for model selection, making 5.56% an optimistic selected estimate. The authors should either use a proper held-out validation set, describe the selection rule explicitly, or present the 5.56% as a best-on-eval number rather tha
  2. [Section IV-B and Table IV] The claim 'Our best result also outperforms other single models reported from recent papers in Table IV' is internally contradicted by the same table: WavLM-RawBoost (FT-DA) is listed with eval EER 5.5%, which is lower than 5.56%. The following sentence hedges: 'Specifically, we achieve comparable performance to a finetuned WavLM model with data augmentation'. This inconsistency affects the central SOTA claim in the abstract and conclusion. The comparison should be restricted to single models without augmentation, or the SOTA wording should be removed.
  3. [Table V, DFADD row] The reported EER of 0.00% on DFADD is suspicious and should be discussed. A perfect EER on a real, non-trivial benchmark is remarkable and suggests either that the evaluation protocol is too easy, that there is a label/feature leakage, or that the test set is too small. Since the cross-dataset generalization is a secondary claim, this does not by itself undermine the main contribution, but the authors should provide the number of trials, error bars, or a discussion of why the result is exactly zero.
minor comments (6)
  1. [Section II-A and References] Reference [23] is 'A survey on LoRa networking'; this is about long-range wireless communication, not Low-Rank Adaptation. The citation for LoRA memory footprint should be the original LoRA paper or a relevant PEFT survey.
  2. [Section IV-B] Typos: 'forzen' should be 'frozen', 'sigle' should be 'single', and 'illutrated' in Section III-B should be 'illustrated'.
  3. [Section IV-C, Table VI caption] 'Tabel' should be 'Table'.
  4. [Equation (10)] The notation NX for the number of samples is undefined and unclear; use N_X to avoid confusion with the number of experts N.
  5. [Table III and Section III-B] The paper should clarify the relationship between M (number of MoLEx modules) and L (number of transformer layers). Table III shows rows where these differ; the text should explain which layers receive MoLEx modules.
  6. [Section IV-B] The phrase 'comparable performance to a finetuned WavLM model with data augmentation' should be integrated into the abstract/conclusion consistently, rather than the standalone 'state-of-the-art' claim.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation in MoLEx reduces to its own inputs; the only author-overlap citation (attentive merging) is a peer-reviewed component with independent evaluation, while the eval-selected 5.56% EER is a generalization risk, not a circular step.

full rationale

The MoLEx construction is self-contained arithmetic: Eq. (2)-(5) define frozen WavLM FFN weights with additive low-rank expert outputs selected by a top-K noisy gating network, Eq. (8) defines an orthogonality regularizer, and Eq. (9) defines the effective-rank diagnostic. The reported EER numbers in Tables III-V are direct evaluations on held-out partitions of ASVSpoof 5 and other corpora; they are not algebraic consequences of the definitions. The orthogonality-loss experiment (Table II) measures rank via SVD after training, so it is an empirical check rather than a circular prediction. The only author-overlap citation is the attentive-merging module in Section II-B, described as 'proposed and evaluated in [21], [26]'; [21] and [26] are prior published works with their own external evaluations, and the module is a component rather than a premise that forces the MoLEx result, so this self-citation is not load-bearing in the circular sense. There is a legitimate concern that the optimal hyperparameters (r=32, K/N=4/12, L=12, M=12) were selected after inspecting eval EERs in Table III, since some dev-lower configurations have higher eval EERs; this affects the unbiasedness of 5.56% as a generalization estimate, but it is a model-selection/data-snooping issue, not an equation that reduces to its own output. The claim that the result 'outperforms other single models' is also internally questionable against Table IV's WavLM-RawBoost 5.5%, but that is a correctness inconsistency, not circularity. Therefore no significant circularity is present; score 2 reflects the minor non-load-bearing self-citation and the otherwise self-contained evaluation chain.

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

The paper introduces no new physical or modelling entity. The method is a composition of known components (LoRA, MoE routing, attentive merging) with an orthogonality penalty. The main caveats are the many hand-chosen hyperparameters and the reliance on the authors' prior attentive merging module.

free parameters (5)
  • LoRA rank r = 32 (best); 8,16,64 also tried
    Chosen by model selection; affects EER from 5.56% (r=32) to 7.51% (r=8).
  • Top-K expert selection = K/N=4/12 (best)
    Varied K/N from 2/6 to 12/12; the chosen value gives best eval EER but not best dev EER.
  • Number of MoLEx modules M = 12 (best)
    M=6 gives EER 7.23 vs 5.56 for M=12.
  • Number of transformer layers L = 12
    L=24 performs worse (15.82%), L=12 selected.
  • Exponent tau for rank threshold = 10^-2, 10^-3
    Used only for the rank analysis in Table II, not for training.
assumptions (4)
  • domain assumption WavLM-Large pre-trained weights remain frozen and provide a strong feature extractor for anti-spoofing
    The whole method relies on the SSL backbone; performance is measured relative to it.
  • domain assumption ASVSpoof 5 dataset labels and partitions are correct
    EERs are computed on the provided dev and eval sets without independent verification.
  • standard math The noisy top-K gating mechanism from Shazeer et al. is implemented as described
    Eq. (3)-(5) use a standard softmax gating; the paper does not detail noise scheduling.
  • domain assumption Attentive merging module from Pan et al. (2024) is effective
    The classifier uses attentive merging from the authors' prior work [21]; performance depends on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoLEx: Mixture of LoRA Experts in Speech Self-Supervised Models for Audio Deepfake Detection." pith.science (2026). https://pith.science/paper/OL3HTKHK

@misc{pith2026250909175,
  author       = {Pith},
  title        = {Pith review of: MoLEx: Mixture of LoRA Experts in Speech Self-Supervised Models for Audio Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OL3HTKHK}},
  note         = {Machine review of arXiv:2509.09175}
}
read the original abstract

While self-supervised learning (SSL)-based models have boosted audio deepfake detection accuracy, fully finetuning them is computationally expensive. To address this, we propose a parameter-efficient framework that combines Low-Rank Adaptation with a Mixture-of-Experts router, called Mixture of LoRA Experts (MoLEx). It preserves pre-trained knowledge of SSL models while efficiently finetuning only selected experts, reducing training costs while maintaining robust performance. The observed utility of experts during inference shows the router reactivates the same experts for similar attacks but switches to other experts for novel spoofs, confirming MoLEx's domain-aware adaptability. MoLEx additionally offers flexibility for domain adaptation by allowing extra experts to be trained without modifying the entire model. We mainly evaluate our approach on the ASVSpoof 5 dataset and achieve the state-of-the-art (SOTA) equal error rate (EER) of 5.56% on the evaluation set without augmentation.

Figures

Figures reproduced from arXiv: 2509.09175 by the authors.

Figure 1
Figure 1. Diagram of the MoLEx in speech SSL model for audio deepfake detection. Grey colored blocks are frozen during [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of extending MoLEx module. Only new [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The utilization scores for the cross-data evaluation experiment in Table V. Each subfigure illustrates the average gating [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Teffic-Audio: Tell Fact from Fiction

    cs.SD 2026-07 conditional novelty 4.0 of 10

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

Reference graph

Works this paper leans on

62 extracted references · 17 linked inside Pith · cited by 1 Pith paper

  1. [11]

    X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with appli- cations in protein mechanics and molecular design,

    E. L. Buehler and M. J. Buehler, “X-lora: Mixture of low-rank adapter experts, a flexible framework for large language models with appli- cations in protein mechanics and molecular design,”APL Machine Learning, vol. 2, no. 2, 2024

    2024

  2. [19]

    Mixture of lora experts,

    X. Wu, S. Huang, and F. Wei, “Mixture of lora experts,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  3. [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

  4. [1]

    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. H. Kinnunen, N. Evans, K. A. Lee, and J. Yamagishi, “Asvspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 1–8

    2024

  5. [2]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models,

    Z. Du, Y . Wang, Q. Chen, X. Shi, X. Lv, T. Zhao, Z. Gao, Y . Yang, C. Gao, H. Wanget al., “Cosyvoice 2: Scalable streaming speech synthesis with large language models,”arXiv preprint arXiv:2412.10117, 2024

    arXiv 2024

  6. [3]

    Aasist3: Kan-enhanced aasist speech deepfake detection using ssl features and additional regularization for the asvspoof 2024 challenge,

    K. Borodin, V . Kudryavtsev, D. Korzh, A. Efimenko, G. Mkrtchian, M. Gorodnichev, and O. Y . Rogov, “Aasist3: Kan-enhanced aasist speech deepfake detection using ssl features and additional regularization for the asvspoof 2024 challenge,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 48– 55

    2024

  7. [4]

    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,” Advances in neural information processing systems, vol. 33, pp. 12 449– 12 460, 2020

    2020

  8. [5]

    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

Show all 62 references
  1. [6]

    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. Xiaoet al., “Wavlm: Large-scale self-supervised pre- training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  2. [7]

    Learn from real: reality defender’s submission to asvspoof5 challenge,

    Y . Zhu, C. Goel, S. Koppisetti, T. Tran, A. Kumar, and G. Bharaj, “Learn from real: reality defender’s submission to asvspoof5 challenge,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 116–123

    2024

  3. [8]

    Peft-ser: On the use of parameter efficient transfer learning approaches for speech emotion recognition using pre- trained speech models,

    T. Feng and S. Narayanan, “Peft-ser: On the use of parameter efficient transfer learning approaches for speech emotion recognition using pre- trained speech models,” in2023 11th International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2023, pp. 1–8

    2023

  4. [9]

    Peft for speech: Unveiling optimal placement, merging strategies, and ensemble techniques,

    T.-H. Lin, H.-S. Wang, H.-Y . Weng, K.-C. Peng, Z.-C. Chen, and H.-y. Lee, “Peft for speech: Unveiling optimal placement, merging strategies, and ensemble techniques,”arXiv preprint arXiv:2401.02122, 2024

    2024 arXiv

  5. [10]

    Adalora: Adaptive budget allocation for parameter-efficient fine-tuning,

    Q. Zhang, M. Chen, A. Bukharin, N. Karampatziakis, P. He, Y . Cheng, W. Chen, and T. Zhao, “Adalora: Adaptive budget allocation for parameter-efficient fine-tuning,”arXiv preprint arXiv:2303.10512, 2023

    2023 arXiv

  6. [12]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  7. [13]

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

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,”arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  8. [14]

    Mixture of a million experts,

    X. O. He, “Mixture of a million experts,”arXiv preprint arXiv:2407.04153, 2024

    2024 arXiv

  9. [15]

    Improving speech emotion recognition by fusing self-supervised learning and spectral features via mixture of experts,

    J. Hyeon, Y .-H. Oh, Y .-J. Lee, and H.-J. Choi, “Improving speech emotion recognition by fusing self-supervised learning and spectral features via mixture of experts,”Data & Knowledge Engineering, vol. 150, p. 102262, 2024

    2024

  10. [16]

    Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models,

    Y . Wang, S. Mukherjee, X. Liu, J. Gao, A. H. Awadallah, and J. Gao, “Adamix: Mixture-of-adapter for parameter-efficient tuning of large language models,”arXiv preprint arXiv:2205.12410, vol. 1, no. 2, p. 4, 2022

    2022 arXiv

  11. [17]

    Adaptermix: Exploring the efficacy of mixture of adapters for low-resource tts adaptation,

    A. Mehrish, A. Ramesh Kashyap, L. Yingting, N. Majumder, and S. Poria, “Adaptermix: Exploring the efficacy of mixture of adapters for low-resource tts adaptation,” inInterspeech 2023, 2023, pp. 4284– 4288

    2023

  12. [18]

    Moe-ffd: Mixture of experts for generalized and parameter-efficient face forgery detection,

    C. Kong, A. Luo, S. Xia, Y . Yu, H. Li, and A. C. Kot, “Moe-ffd: Mixture of experts for generalized and parameter-efficient face forgery detection,” CoRR, 2024

    2024

  13. [20]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  14. [22]

    Scaling language models: Methods, analysis & insights from training gopher,

    J. W. Rae, S. Borgeaud, T. Cai, K. Millican, J. Hoffmann, F. Song, J. Aslanides, S. Henderson, R. Ring, S. Younget al., “Scaling language models: Methods, analysis & insights from training gopher,”arXiv preprint arXiv:2112.11446, 2021

    2021 arXiv

  15. [23]

    A survey on lora networking: Research problems, current solutions, and open issues,

    J. P. S. Sundaram, W. Du, and Z. Zhao, “A survey on lora networking: Research problems, current solutions, and open issues,”IEEE Commu- nications Surveys & Tutorials, vol. 22, no. 1, pp. 371–388, 2019

    2019

  16. [24]

    Mixture-of-experts with expert choice routing,

    Y . Zhou, T. Lei, H. Liu, N. Du, Y . Huang, V . Zhao, A. M. Dai, Q. V . Le, J. Laudonet al., “Mixture-of-experts with expert choice routing,” Advances in Neural Information Processing Systems, vol. 35, pp. 7103– 7114, 2022

    2022

  17. [25]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,

    W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,”Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022

    2022

  18. [26]

    Speech foundation model ensembles for the controlled singing voice deepfake detection (ctrsvdd) challenge 2024,

    A. Guragain, T. Liu, Z. Pan, H. B. Sailor, and Q. Wang, “Speech foundation model ensembles for the controlled singing voice deepfake detection (ctrsvdd) challenge 2024,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 774–781

    2024

  19. [27]

    The expressive power of low-rank adaptation,

    Y . Zeng and K. Lee, “The expressive power of low-rank adaptation,” arXiv preprint arXiv:2310.17513, 2023

    2023 arXiv

  20. [28]

    ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale,

    ASVspoof 5 Challenge, “ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale,” 2024, accessed: Feb 11, 2025. [Online]. Available: https://codalab.lisn.upsaclay.fr/competitions/19383#results

    2024

  21. [29]

    The Singular Value Decomposition (SVD) and Low-Rank Matrix Approximations,

    Stanford University, “The Singular Value Decomposition (SVD) and Low-Rank Matrix Approximations,” 2024, accessed: Feb 11, 2025. [Online]. Available: https://web.stanford.edu/class/cs168/l/l9.pdf

    2024

  22. [30]

    A single end-to-end voice anti-spoofing model with graph attention and feature aggregation for asvspoof 5 challenge,

    W. Xia, H. Peng, L. Li, and Y . Ren, “A single end-to-end voice anti-spoofing model with graph attention and feature aggregation for asvspoof 5 challenge,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 124–130

    2024

  23. [31]

    A study of guided masking data augmentation for deepfake speech detection,

    D.-T. Truong, Y . Wang, K. A. Lee, M. Li, H. Nishizaki, and E. S. Chng, “A study of guided masking data augmentation for deepfake speech detection,” inThe Automatic Speaker Verification Spoofing Countermea- sures Workshop (ASVspoof 2024), 2024, pp. 176–180

    2024

  24. [32]

    Enhancing spoofing detection in asvspoof 5 workshop 2024: fusion of wavlm-resnet18-sa for optimal performance against speech deepfakes,

    P.-C. Chan, W.-Y . Chen, and J.-C. Wang, “Enhancing spoofing detection in asvspoof 5 workshop 2024: fusion of wavlm-resnet18-sa for optimal performance against speech deepfakes,” inThe Automatic Speaker Ver- ification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp...

    2024

  25. [33]

    Exploring wavlm back- ends for speech spoofing and deepfake detection,

    T. Stourbe, V . Miara, T. Lepage, and R. Dehak, “Exploring wavlm back- ends for speech spoofing and deepfake detection,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 72–78

    2024

  26. [34]

    Wavlm model ensemble for audio deepfake detection,

    D. Combei, A. Stan, D. Oneata, and H. Cucu, “Wavlm model ensemble for audio deepfake detection,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 170– 175

    2024

  27. [35]

    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

  28. [36]

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

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kinnunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautschet al., “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 25...

    2021

  29. [37]

    Does audio deepfake detection generalize?

    N. M ¨uller, P. Czempin, F. Diekmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?” inInterspeech 2022, 2022, pp. 2783–2787

    2022

  30. [38]

    Robust audio deepfake detection: Exploring front-/back-end combinations and data augmentation strate- gies for the asvspoof5 challenge,

    K. Sch ¨afer, M. Neu, and J.-E. Choi, “Robust audio deepfake detection: Exploring front-/back-end combinations and data augmentation strate- gies for the asvspoof5 challenge,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop 2024, 2024

    2024

  31. [39]

    Ai-synthesized voice detection using neural vocoder artifacts,

    C. Sun, S. Jia, S. Hou, and S. Lyu, “Ai-synthesized voice detection using neural vocoder artifacts,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 904–912

    2023

  32. [40]

    Dfadd: The diffusion and flow-matching based audio deepfake dataset,

    J. Du, I.-M. Lin, I.-H. Chiu, X. Chen, H. Wu, W. Ren, Y . Tsao, H.-y. Lee, and J.-S. R. Jang, “Dfadd: The diffusion and flow-matching based audio deepfake dataset,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 921–928

    2024

  33. [41]

    The Fake-or-Real Dataset,

    APTLY lab, “The Fake-or-Real Dataset,” https://bil.eecs.yorku.ca/datasets/, 2025

    2025

  34. [42]

    Speech arena: Speech deepfake leaderboard,

    S. Arena, “Speech arena: Speech deepfake leaderboard,” https://huggingface.co/spaces/Speech-Arena-2025/Speech-DF-Arena, 2025

    2025

  35. [43]

    Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit,

    C. Veaux, J. Yamagishi, and K. MacDonald, “Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit,” 2017

    2017

  36. [44]

    Wavenet: A generative model for raw audio,

    A. Van Den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, K. Kavukcuogluet al., “Wavenet: A generative model for raw audio,”arXiv preprint arXiv:1609.03499, vol. 12, 2016

    2016 arXiv

  37. [45]

    Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,

    J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y . Zhang, Y . Wang, R. Skerrv-Ryanet al., “Natural tts synthesis by conditioning wavenet on mel spectrogram predictions,” in2018 IEEE international conference on acoustics, speech and signal processing (...

    2018

  38. [46]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,” inInterspeech 2020, 2020, pp. 2757–2761

    2020

  39. [47]

    Diffusion-based voice conversion with fast maximum likelihood sam- pling scheme,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, M. Kudinov, and J. Wei, “Diffusion-based voice conversion with fast maximum likelihood sam- pling scheme,”arXiv preprint arXiv:2109.13821, 2021

    2021 arXiv

  40. [48]

    Low-resource multilingual and zero-shot multispeaker tts,

    F. Lux, J. Koch, and N. T. Vu, “Low-resource multilingual and zero-shot multispeaker tts,”arXiv preprint arXiv:2210.12223, 2022

    2022 arXiv

  41. [49]

    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,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 5530–5540

    2021

  42. [50]

    Glow-tts: A generative flow for text-to-speech via monotonic alignment search,

    J. Kim, S. Kim, J. Kong, and S. Yoon, “Glow-tts: A generative flow for text-to-speech via monotonic alignment search,”Advances in Neural Information Processing Systems, vol. 33, 2020

    2020

  43. [51]

    Creating new language and voice compo- nents for the updated marytts text-to-speech synthesis platform,

    I. Steiner and S. L. Maguer, “Creating new language and voice compo- nents for the updated marytts text-to-speech synthesis platform,”arXiv preprint arXiv:1712.04787, 2017

    2017 arXiv

  44. [52]

    P-flow: A fast and data-efficient zero-shot tts through speech prompting,

    S. Kim, K. Shih, J. F. Santos, E. Bakhturina, M. Desta, R. Valle, S. Yoon, B. Catanzaroet al., “P-flow: A fast and data-efficient zero-shot tts through speech prompting,”Advances in Neural Information Processing Systems, vol. 36, pp. 74 213–74 228, 2023

    2023

  45. [53]

    Matcha-tts: A fast tts architecture with conditional flow matching,

    S. Mehta, R. Tu, J. Beskow, ´E. Sz ´ekely, and G. E. Henter, “Matcha-tts: A fast tts architecture with conditional flow matching,” inICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 341–11 345

    2024

  46. [54]

    Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,

    Y . A. Li, C. Han, V . Raghavan, G. Mischler, and N. Mesgarani, “Styletts 2: Towards human-level text-to-speech through style diffusion and adversarial training with large speech language models,”Advances in Neural Information Processing Systems, vol. 36, pp. 19 594–19 621, 2023

    2023

  47. [55]

    Grad- tts: A diffusion probabilistic model for text-to-speech,

    V . Popov, I. V ovk, V . Gogoryan, T. Sadekova, and M. Kudinov, “Grad- tts: A diffusion probabilistic model for text-to-speech,” inInternational conference on machine learning. PMLR, 2021, pp. 8599–8608

    2021

  48. [56]

    Libritts: A corpus derived from librispeech for text-to-speech,

    H. Zen, V . Dang, R. Clark, Y . Zhang, R. J. Weiss, Y . Jia, Z. Chen, and Y . Wu, “Libritts: A corpus derived from librispeech for text-to-speech,” arXiv preprint arXiv:1904.02882, 2019

    1904 arXiv

  49. [57]

    Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram,

    R. Yamamoto, E. Song, and J.-M. Kim, “Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram,” inICASSP 2020-2020 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2...

    2020

  50. [58]

    Dif- fwave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Dif- fwave: A versatile diffusion model for audio synthesis,”arXiv preprint arXiv:2009.09761, 2020

    2009 arXiv

  51. [59]

    Wavegrad: Estimating gradients for waveform generation,

    N. Chen, Y . Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating gradients for waveform generation,”arXiv preprint arXiv:2009.00713, 2020

    2009 arXiv

  52. [60]

    V oxForge speech dataset,

    , “V oxForge speech dataset,” https://www.voxforge.org/, 2025

    2025

  53. [61]

    Deep voice 3: Scaling text-to-speech with convolutional sequence learning,

    W. Ping, K. Peng, A. Gibiansky, S. O. Arik, A. Kannan, S. Narang, J. Raiman, and J. Miller, “Deep voice 3: Scaling text-to-speech with convolutional sequence learning,”arXiv preprint arXiv:1710.07654, 2017

    2017 arXiv

  54. [62]

    Catastrophic forgetting, rehearsal and pseudorehearsal,

    A. Robins, “Catastrophic forgetting, rehearsal and pseudorehearsal,” Connection Science, vol. 7, no. 2, pp. 123–146, 1995

    1995

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.