pith. sign in

arxiv: 2606.16532 · v2 · pith:ZL3UMUTJnew · submitted 2026-06-15 · 💻 cs.SD · cs.AI

Dual-Granularity Orthogonal Disentanglement for Generalizable Audio Deepfake Detection

Zhuodong Liu , Hugen Lv , Xiangyu Li , Chunhong Yuan This is my paper

Pith reviewed 2026-06-27 03:08 UTC · model grok-4.3

classification 💻 cs.SD cs.AI
keywords audio deepfake detectionorthogonal disentanglementspeaker identity leakagecross-dataset generalizationcosine orthogonalitycross-covariance regularizationcurriculum learning
0
0 comments X

The pith

Enforcing sample-level and batch-level orthogonality disentangles speaker identity from synthesis artifacts in audio deepfake detectors.

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

The paper aims to fix the tendency of audio deepfake detectors to latch onto speaker identity instead of the actual synthesis traces that indicate fakes. It does so by adding two orthogonal constraints: one that makes individual sample embeddings point in unrelated directions via cosine similarity, and another that removes linear correlations across dimensions within each training batch. These constraints are applied through a curriculum that ramps up their strength over training epochs, avoiding any need for extra networks or adversarial components. If the approach works, detectors should transfer more reliably to new speakers and new synthesis methods while keeping detection accuracy intact.

Core claim

The central claim is that a dual-granularity orthogonal disentanglement framework, consisting of sample-level cosine orthogonality and batch-level cross-covariance regularization applied under a progressive curriculum schedule, removes implicit speaker-identity leakage from the learned embeddings while preserving synthesis-artifact cues, thereby improving cross-dataset generalization in audio deepfake detection without auxiliary networks or adversarial training dynamics.

What carries the argument

Dual-granularity orthogonal disentanglement, which combines per-sample cosine orthogonality for directional decorrelation with per-batch cross-covariance regularization to eliminate linear correlations, scheduled by a curriculum that gradually strengthens the constraints.

If this is right

  • The method reports equal error rates of 1.35 percent on ASVspoof 2019 LA, 7.88 percent on ASVspoof 2021 DF, and 21.58 percent on In-the-Wild data.
  • It improves absolute cross-dataset transfer by 2.60 percent over gradient reversal disentanglement baselines.
  • No auxiliary networks or adversarial losses are required to achieve the reported separation of identity and artifact features.
  • The curriculum schedule allows the orthogonality constraints to be introduced without destabilizing early training.

Where Pith is reading between the lines

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

  • The same two-level orthogonality pattern could be tested on image or video deepfake detectors where identity leakage is also observed.
  • If the batch-level covariance term proves critical, it suggests that linear decorrelation across dimensions is a practical surrogate for full statistical independence in embedding spaces.
  • The curriculum mechanism might transfer to other regularization tasks that need to avoid early over-constraining of the model.

Load-bearing premise

Enforcing orthogonality between embeddings will separate speaker identity information from synthesis artifact cues without discarding information needed for accurate detection.

What would settle it

Train the model on data where speaker identity is deliberately made predictive of the fake label, then measure whether equal error rate rises sharply on a test set where speaker identity and label are uncorrelated.

Figures

Figures reproduced from arXiv: 2606.16532 by Chunhong Yuan, Hugen Lv, Xiangyu Li, Zhuodong Liu.

Figure 1
Figure 1. Figure 1: Overview of the proposed dual-branch architecture with dual-granularity orthogonal disentanglement. nality constraints, preventing premature feature collapse while achieving stronger final disentanglement. Third, we demon￾strate through comprehensive experiments that this approach achieves 1.35% and 7.88% EER on ASVspoof 2019 LA and 2021 DF [34], respectively, and 21.58% EER on In-the-Wild, performance com… view at source ↗
Figure 2
Figure 2. Figure 2: shows the sensitivity to βmax on ASVspoof 2021 DF. Without orthogonality (βmax = 0), the average cosine similar￾ity between embeddings reaches 0.342, indicating substantial information overlap between the two branches. As βmax in￾creases, cosine similarity decreases monotonically, confirming effective orthogonality enforcement. Notably, dual-granularity disentanglement with curriculum scheduling broadens t… view at source ↗
Figure 3
Figure 3. Figure 3: t-SNE visualization of content embeddings zc on In￾the-Wild [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

Audio deepfake detectors often fail to generalize across speakers, as they learn speaker-identity features rather than synthesis artifacts, known as implicit identity leakage. Existing methods address this but incur architectural complexity or training instability. This paper proposes a dual-granularity orthogonal disentanglement framework enforcing feature independence at two levels: sample-level cosine orthogonality captures directional decorrelation, while batch-level cross-covariance regularization eliminates linear correlations across embedding dimensions. A curriculum disentanglement schedule progressively strengthens the orthogonality constraint without auxiliary networks or adversarial dynamics. Experiments on ASVspoof 2019 LA, ASVspoof 2021 DF, and In-the-Wild datasets demonstrate that the proposed method achieves 1.35%, 7.88%, and 21.58% equal error rates (EER), respectively, surpassing gradient reversal disentanglement by 2.60% absolute on cross-dataset transfer.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a dual-granularity orthogonal disentanglement framework for audio deepfake detection to address implicit identity leakage. It enforces feature independence via sample-level cosine orthogonality and batch-level cross-covariance regularization, combined with a curriculum disentanglement schedule that progressively strengthens the constraint without auxiliary networks or adversarial training. Experiments report EERs of 1.35% on ASVspoof 2019 LA, 7.88% on ASVspoof 2021 DF, and 21.58% on In-the-Wild, with a 2.60% absolute improvement over gradient reversal disentanglement on cross-dataset transfer.

Significance. If the orthogonality terms demonstrably separate speaker identity from synthesis artifacts, the approach would provide a simpler, more stable alternative to adversarial disentanglement methods for improving generalization in audio deepfake detection. The curriculum schedule is a constructive design element that may aid training stability, and the multi-dataset EER reporting supplies concrete benchmarks.

major comments (2)
  1. [Abstract and Experiments] Abstract and Experiments: The central claim that sample-level cosine orthogonality plus batch-level cross-covariance regularization removes speaker-identity information while retaining synthesis-artifact cues is load-bearing for the reported EER gains, yet the manuscript supplies only downstream EER numbers and the 2.60% cross-dataset improvement. No auxiliary metrics (speaker-ID accuracy on embeddings, mutual information estimates, or controlled ablations isolating the orthogonality terms) are presented to confirm the intended separation occurred rather than generic regularization effects.
  2. [Method] Method (curriculum disentanglement schedule): The progressive strengthening of the orthogonality constraint is described as key to avoiding training instability, but no sensitivity analysis, ablation on schedule hyperparameters, or examination of the resulting identity-artifact trade-off is provided. This directly affects interpretability of the cross-dataset EER results.
minor comments (1)
  1. [Method] Notation in the batch-level cross-covariance term should be expanded with an explicit equation to support reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments below and will incorporate revisions to strengthen the evidence for our claims.

read point-by-point responses

  1. Referee: [Abstract and Experiments] Abstract and Experiments: The central claim that sample-level cosine orthogonality plus batch-level cross-covariance regularization removes speaker-identity information while retaining synthesis-artifact cues is load-bearing for the reported EER gains, yet the manuscript supplies only downstream EER numbers and the 2.60% cross-dataset improvement. No auxiliary metrics (speaker-ID accuracy on embeddings, mutual information estimates, or controlled ablations isolating the orthogonality terms) are presented to confirm the intended separation occurred rather than generic regularization effects.

    Authors: We agree that auxiliary metrics would provide more direct confirmation that the orthogonality terms achieve the intended separation rather than acting as generic regularization. The 2.60% cross-dataset gain over gradient reversal (which uses a comparable adversarial mechanism) offers indirect support, but this does not fully isolate the contribution of our dual-granularity terms. In the revised manuscript we will add speaker identification accuracy on the learned embeddings and controlled ablations that isolate the sample-level cosine and batch-level cross-covariance terms. revision: yes

  2. Referee: [Method] Method (curriculum disentanglement schedule): The progressive strengthening of the orthogonality constraint is described as key to avoiding training instability, but no sensitivity analysis, ablation on schedule hyperparameters, or examination of the resulting identity-artifact trade-off is provided. This directly affects interpretability of the cross-dataset EER results.

    Authors: The curriculum schedule was introduced to gradually ramp up the orthogonality constraint and thereby improve training stability without auxiliary networks. We acknowledge that the original submission lacks sensitivity analysis on its hyperparameters and does not quantify any identity-artifact trade-off. In revision we will add ablations varying the starting epoch and ramp rate, together with the resulting EERs and any observed changes in embedding properties. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results on external benchmarks are independent of internal definitions

full rationale

The paper introduces a dual-granularity orthogonal disentanglement method with sample-level cosine orthogonality and batch-level cross-covariance regularization, trained via a curriculum schedule, then reports EER numbers on the external ASVspoof 2019 LA, ASVspoof 2021 DF, and In-the-Wild datasets. These performance figures are obtained by standard supervised training and evaluation on public benchmarks; they do not reduce to any fitted parameter that is later renamed as a prediction, nor to any self-citation chain that supplies the uniqueness or correctness of the orthogonality constraints. No equations or claims in the provided text equate the target separation of speaker identity from synthesis artifacts to the orthogonality terms by construction. The derivation chain therefore remains self-contained against external data rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no equations, sections, or implementation details provided to identify free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5685 in / 1127 out tokens · 31533 ms · 2026-06-27T03:08:04.584450+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    Introduction Advances in voice conversion [1, 2] and text-to-speech [3, 4] have enabled highly realistic synthetic speech, threatening speaker verification systems and enabling fraud and misinfor- mation. While state-of-the-art detectors achieve below 2% equal error rates (EER) on ASVspoof 2019 LA [5, 6, 7], per- formance degrades to over 20% EER on real-...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  2. [2]

    Problem Formulation LetX∈R F×T denote the log-mel spectrogram of an input utterance withFmel frequency bins andTtime frames

    Proposed Method 2.1. Problem Formulation LetX∈R F×T denote the log-mel spectrogram of an input utterance withFmel frequency bins andTtime frames. Given a labeled training setD={(X i, yi, si)}N i=1, wherey i ∈ {0,1} indicates spoofed or bonafide ands i ∈ {1, . . . , K}denotes speaker identity, the goal of this paper is to learn a detector that generalizes ...

  3. [3]

    The training set contains 22,617 bonafide and 22,296 spoofed utterances from 107 speakers

    Experimental Setup We evaluate on ASVspoof 2021 DF [34], which contains bonafide speech and spoofed samples from over 100 differ- ent synthesis systems including VCC2018 and VCC2020 voice conversion submissions. The training set contains 22,617 bonafide and 22,296 spoofed utterances from 107 speakers. For cross-dataset evaluation, we test models trained o...

    2021

  4. [4]

    In-Domain Detection Table 1 presents detection performance on ASVspoof 2019 LA and 2021 DF

    Results and Discussion 4.1. In-Domain Detection Table 1 presents detection performance on ASVspoof 2019 LA and 2021 DF. On ASVspoof 2019 LA, the proposed method achieves 1.35% EER, ranking second among non-pretrained methods behind AASIST (0.83%), which benefits from graph- based spectro-temporal modeling optimized for in-domain con- ditions. The proposed...

    2019

  5. [5]

    Conclusion This paper presented a dual-granularity orthogonal disentangle- ment framework combining sample-level cosine orthogonality with batch-level cross-covariance regularization under a cur- riculum schedule, enforcing speaker-artifact separation with- out auxiliary networks or adversarial training. This lightweight approach (2.1M parameters) achieve...

  6. [6]

    All scientific content, experimental design, implementa- tion, analysis, and conclusions are the sole work of the authors

    Generative AI Use Disclosure Generative AI tools were used for English language polish- ing and LATEX formatting assistance during manuscript prepa- ration. All scientific content, experimental design, implementa- tion, analysis, and conclusions are the sole work of the authors

  7. [7]

    An overview of voice conversion and its challenges: From statistical modeling to deep learning,

    B. Sisman, J. Yamagishi, S. King, and H. Li, “An overview of voice conversion and its challenges: From statistical modeling to deep learning,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 132–157, Nov. 2021

    2021

  8. [8]

    Any- to-many voice conversion with location-relative sequence-to- sequence modeling,

    S. Liu, Y . Cao, D. Wang, X. Wu, X. Liu, and H. Meng, “Any- to-many voice conversion with location-relative sequence-to- sequence modeling,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1717–1728, Apr. 2021

    2021

  9. [9]

    Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,

    E. Casanovaet al., “Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,” inInternational Conference on Machine Learning, Baltimore, MD, USA, 2022, pp. 2709–2720

    2022

  10. [10]

    Fastspeech 2: Fast and high-quality end-to-end text to speech,

    Y . Renet al., “Fastspeech 2: Fast and high-quality end-to-end text to speech,” inInternational Conference on Learning Representa- tions, Virtual, May 2021

    2021

  11. [11]

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

    J.-w. Junget al., “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 6367–6371

    2022

  12. [12]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inIEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing, Toronto, ON, Canada, 2021, pp. 6369–6373

    2021

  13. [13]

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

    X. Wanget al., “Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech,”Computer Speech & Language, vol. 64, p. 101114, Nov. 2020

    2019

  14. [14]

    Does audio deepfake detection generalize?

    N. M. M ¨ulleret al., “Does audio deepfake detection generalize?” inInterspeech, Incheon, Korea, 2022, pp. 2783–2787

    2022

  15. [15]

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

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

    2024

  16. [16]

    Beyond identity: A generalizable approach for deepfake audio detection,

    Y . Ahmadiadli, X.-P. Zhang, and N. M. Khan, “Beyond identity: A generalizable approach for deepfake audio detection,” 2025, [Online]. Available: https://arxiv.org/abs/2505.06766

    work page arXiv 2025

  17. [17]

    Audio deepfake detection: A survey,

    J. Yi, C. Wang, J. Tao, X. Zhang, C. Y . Zhang, and Y . Zhao, “Audio deepfake detection: A survey,” 2023, [Online]. Available: https://arxiv.org/abs/2308.14970

    work page arXiv 2023

  18. [18]

    Spoofing-aware speaker verification with un- supervised domain adaptation,

    X. Wanget al., “Spoofing-aware speaker verification with un- supervised domain adaptation,” inIEEE International Confer- ence on Acoustics, Speech and Signal Processing, Rhodes Island, Greece, 2023, pp. 1–5

    2023

  19. [19]

    Domain generalization via aggregation and separation for audio deepfake detection,

    Y . Xie, H. Cheng, Y . Wang, and L. Ye, “Domain generalization via aggregation and separation for audio deepfake detection,”IEEE Transactions on Information Forensics and Security, vol. 19, pp. 344–358, 2024

    2024

  20. [20]

    ASVspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detec- tion using crowdsourced speech,

    X. Wanget al., “ASVspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detec- tion using crowdsourced speech,”Computer Speech & Language, vol. 95, p. 101825, Jan. 2026

    2026

  21. [21]

    A comparison of features for synthetic speech detection,

    M. Sahidullah, T. Kinnunen, and C. Hanilci, “A comparison of features for synthetic speech detection,” inInterspeech, Dresden, Germany, 2015, pp. 2087–2091

    2015

  22. [22]

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

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” inAdvances in Neural Information Processing Systems, vol. 33, Virtual, Dec. 2020, pp. 12 449–12 460

    2020

  23. [23]

    Vicomtech audio deepfake detection system based on wav2vec 2.0 for the 2022 add chal- lenge,

    J. M. Martin-Donas and A. Alvarez, “Vicomtech audio deepfake detection system based on wav2vec 2.0 for the 2022 add chal- lenge,” inIEEE International Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 9241–9245

    2022

  24. [24]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Taket al., “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,” inThe Speaker and Language Recognition Workshop (Odyssey), Beijing, China, 2022, pp. 112–119

    2022

  25. [25]

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

    Z. Panet al., “Attentive merging of hidden embeddings from pre- trained speech model for anti-spoofing detection,” inInterspeech, Kos Island, Greece, 2024, pp. 2090–2094

    2024

  26. [26]

    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,”IEEE Transactions on Information Foren- sics and Security, vol. 20, pp. 12 005–12 018, Oct. 2025

    2025

  27. [27]

    X-vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-vectors: Robust dnn embeddings for speaker recognition,” inIEEE International Conference on Acoustics, Speech and Sig- nal Processing, Calgary, AB, Canada, 2018, pp. 5329–5333

    2018

  28. [28]

    Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inInterspeech, Shanghai, China, 2020, pp. 3830–3834

    2020

  29. [29]

    Un- masking real-world audio deepfakes: A data-centric approach,

    D. Combei, A. Stan, D. Oneat ¸ ˘a, N. M ¨uller, and H. Cucu, “Un- masking real-world audio deepfakes: A data-centric approach,” in Interspeech, Rotterdam, Netherlands, Aug. 2025, pp. 5343–5347

    2025

  30. [30]

    Domain- adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. Marchand, and V . Lempitsky, “Domain- adversarial training of neural networks,”Journal of Machine Learning Research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  31. [31]

    Representation learning: A review and new perspectives,

    Y . Bengio, A. Courville, and P. Vincent, “Representation learning: A review and new perspectives,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 8, pp. 1798–1828, Aug. 2013

    2013

  32. [32]

    Alden: Dual-level disentanglement with meta- learning for generalizable audio deepfake detection,

    Y . Xuet al., “Alden: Dual-level disentanglement with meta- learning for generalizable audio deepfake detection,” inProceed- ings of the 33rd ACM International Conference on Multimedia, ser. MM ’25. New York, NY , USA: Association for Computing Machinery, oct 2025, pp. 7277–7286

    2025

  33. [33]

    Safeear: Con- tent privacy-preserving audio deepfake detection,

    X. Li, K. Li, Y . Zheng, C. Yan, X. Ji, and W. Xu, “Safeear: Con- tent privacy-preserving audio deepfake detection,” inACM Con- ference on Computer and Communications Security, Salt Lake City, UT, USA, 2024, pp. 3585–3599

    2024

  34. [34]

    Towards the next frontier in speech representation learning using disentanglement,

    V . Krishna and S. Ganapathy, “Towards the next frontier in speech representation learning using disentanglement,” 2024, [Online]. Available: https://arxiv.org/abs/2407.02543

    work page arXiv 2024

  35. [35]

    ContentVec: An improved self-supervised speech representation by disentangling speakers,

    K. Qianet al., “ContentVec: An improved self-supervised speech representation by disentangling speakers,” inInternational Con- ference on Machine Learning, Baltimore, MD, USA, 2022, pp. 18 003–18 017

    2022

  36. [36]

    Speaker anonymization using orthogonal Householder neural network,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Speaker anonymization using orthogonal Householder neural network,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 3681–3695, Sep. 2023

    2023

  37. [37]

    Speech emo- tion recognition with co-attention based multi-level acoustic infor- mation,

    H. Zou, Y . Si, C. Chen, D. Rajan, and E. S. Chng, “Speech emo- tion recognition with co-attention based multi-level acoustic infor- mation,” inIEEE International Conference on Acoustics, Speech and Signal Processing, Singapore, 2022, pp. 7367–7371

    2022

  38. [38]

    Disentangling spoof trace for generic face anti-spoofing,

    Y . Liu, J. Stehouwer, A. Jourabloo, and X. Liu, “Disentangling spoof trace for generic face anti-spoofing,” inIEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, Seattle, W A, USA, 2020, pp. 8765–8775

    2020

  39. [39]

    Barlow twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow twins: Self-supervised learning via redundancy reduction,” inIn- ternational Conference on Machine Learning, Virtual, Jul. 2021, pp. 12 310–12 320

    2021

  40. [40]

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

    J. Yamagishiet al., “ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection,” inProc. 2021 Edition of the Automatic Speaker Verification and Spoofing Countermea- sures Challenge, 2021, pp. 47–54

    2021

  41. [41]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 2019, pp. 4690–4699

    2019

  42. [42]

    Stc antispoofing systems for the asvspoof 2019 challenge,

    G. Lavrentyevaet al., “Stc antispoofing systems for the asvspoof 2019 challenge,” inInterspeech, Graz, Austria, 2019, pp. 1033– 1037

    2019