Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A fused XLSR-Conformer system with N-pair metric loss traces audio deepfakes to their generating system at 95.61% accuracy in-domain and 44.82% on unseen sources, beating the Wav2Vec2-AASIST baseline by 12 and 18 points respectively.

desk verdict In-domain gains look real, but the OOD claim rests on an unspecified threshold and a single run; the evaluation table also needs cleanup before the paper is publishable. read the letter →

arxiv 2506.02085 v1 pith:ZAVT3CG7 submitted 2025-06-02 cs.SD cs.AIcs.CLeess.AS

Ajinkya Kulkarni , Sandipana Dowerah , Tanel Alumae , Mathew Magimai.-Doss This is my paper
classification cs.SDcs.AIcs.CLeess.AS
keywords audiosourcetracingdeepfakedetectiondeepmetriclearningN-pairlossConformerout-of-domainensemblefusionFrechetDistance
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 tracing an audio deepfake to the text-to-speech or voice-conversion system that produced it is substantially improved by a specific combination: self-supervised Wav2Vec2-XLSR features, a Conformer or state-space back end, a two-stage training recipe that first emphasizes real speech and then disperses fake-sounding embeddings, and a multi-class N-pair metric loss. On the MLAAD source-tracing benchmark, the best fused system reaches 95.61% in-domain accuracy, more than 12 points above the Wav2Vec2-AASIST baseline, and the best single system reaches 44.82% accuracy on unseen deepfake sources versus 26.51% for the baseline. If the combination is correct, it gives forensic tools a practical route to identifying which generator produced a deepfake, including generators not seen during training.

What carries the argument

The load-bearing mechanism is a two-stage training objective, started with binary Real Emphasis (real vs. fake) and continued with Fake Dispersion, where RegMixup is combined with a multi-class N-pair loss that pushes each embedding away from N-1 negative source classes while pulling positives together. The front end is the last hidden layer of Wav2Vec2-XLSR, fed into Conformer, MAMBA, or HYDRA back ends; at inference, Novel Similarity Detection uses cosine similarity plus confidence scaling and a development-set threshold to label samples from unseen generators as out-of-domain. An ensemble fusion that concatenates embeddings and averages logits combines the strengths of selected systems, and Frechet Distance between in-domain and OOD embedding distributions is used as a measure of feature stability.

What would settle it

Re-run the same trained systems across a grid of NSD thresholds and multiple seeds; if the HYDRA system's 44.82% out-of-domain accuracy does not consistently beat the baseline's 26.51% and the Conformer-N-pair system's 38.89% at reasonable cutoffs, the out-of-domain ordering in the paper is not robust to the threshold choice.

Watch

Extended reading notes

Core claim

The paper's central claim, on its own terms, is that a Conformer-based audio source tracer trained with Real Emphasis, Fake Dispersion, and the multi-class N-pair objective outperforms the Wav2Vec2-AASIST baseline in both controlled and open-set tracing. The highest-scoring ensemble, E1 fusing S2 and S3, reports 95.61% accuracy and 95.76% F1 on the in-domain MLAAD evaluation set, against 83.39% and 83.34% for the baseline. In out-of-domain evaluation, the HYDRA-based system S5 reports 44.82% accuracy and 49.31% F1 versus 26.51% and 30.15% for the baseline, while S3 gives the lowest Frechet Distance (6.93), which the paper interprets as stable and generalizable feature representations. The authors conclude that no single system dominates, and that ensemble score-embedding fusion gives the best trade-off between in-domain accuracy and OOD robustness.

Load-bearing premise

The out-of-domain conclusions rely on the Novel Similarity Detection threshold, which is chosen on the development set but neither specified nor varied in the paper, so a different cutoff could reorder the systems.

Editorial extensions

If this is right

  • Adding the multi-class N-pair loss to the RE/FD recipe lifts single-system in-domain accuracy from 83.39% to 95.09% for the Conformer-based S3, a direct reported result.
  • Ensemble fusion of two Conformer variants (S2, S3) pushes in-domain performance to 95.61% while keeping OOD accuracy near S3, making score-embedding fusion the recommended choice when closed-set accuracy matters.
  • The HYDRA-based S5 is preferable when unseen deepfake algorithms are expected, since it records the highest OOD accuracy and lowest calibration error among all systems.
  • Frechet Distance can serve as a model-selection signal: S3 has the lowest value and the most stable features across domains, even though it does not lead in OOD accuracy.
  • The in-domain versus OOD trade-off observed across systems implies that practical deployment should choose between Conformer ensembles for controlled settings and HYDRA for open-set settings.

Reading between the lines

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

  • The finding that low Frechet Distance accompanies stable cross-domain features suggests the metric could be used for model selection before any OOD labels are available, since it requires only the embedding distributions.
  • The N-pair loss component is generic enough to be grafted onto other self-supervised front ends, offering a test of whether the reported gains come from the loss or from the XLSR features specifically.
  • Because the out-of-domain numbers depend on the NSD threshold, re-running the evaluation across a grid of thresholds and seeds would show whether S5's lead over S3 and the baseline is a property of the model or an artifact of the cutoff.
  • The Frechet analysis implicitly assumes approximately Gaussian embedding distributions; a distribution-free alternative might change which model is judged most stable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an audio deepfake source tracing system built on Wav2Vec2-XLSR features with three back-end architectures (Conformer, MAMBA, HYDRA), a two-stage Real Emphasis / Fake Dispersion training framework, an added multi-class N-pair loss, and ensemble score-embedding fusion. The authors evaluate on the MLAAD source tracing benchmark, reporting in-domain accuracy of 95.61% for the best ensemble (E1, S2+S3) versus 83.39% for the Wav2Vec2-AASIST baseline, and out-of-domain accuracy of 44.82% for the HYDRA-only system S5 versus 26.51% for the baseline. The paper also reports Frechet Distance, EER, NLL, and ECE metrics to support claims of feature stability and calibration.

Significance. If the claims were fully supported, the paper would provide a useful, practical result: an open-set audio source tracing system that improves on a published baseline on a public benchmark (MLAAD) while demonstrating a concrete in-domain/OOD trade-off across back-end architectures. The authors are transparent about adopting the RE/FD framework and NSD detector from Xie et al., and the baseline implementation is publicly available, which aids reproducibility. However, the paper's evaluation has several gaps: every result comes from a single run, the out-of-domain decision threshold is unspecified, the reported NLL values are mathematically impossible, and the Frechet Distance analysis relies on an unstated Gaussian assumption. These issues currently prevent the paper from supporting its headline claims at the standard expected for the field.

major comments (4)
  1. [Section 2.3, Table 1] The entire out-of-domain evaluation depends on the Novel Similarity Detection threshold, but the paper never specifies how the development-set threshold is chosen, what selection criterion is used, or what threshold value results. The OOD numbers in Table 1 (e.g., S5 accuracy 44.82% vs. baseline 26.51%) are therefore not reproducible and could change under different legitimate threshold choices. Please specify the threshold selection rule, report the threshold value, and provide a sensitivity analysis over threshold values or a small set of principled alternatives. Additionally, report results across multiple random seeds with means and standard deviations, since all cell values in Table 1 come from a single run.
  2. [Section 3.3, Table 1, Eq. (1)] The NLL values in Table 1 are negative (e.g., -0.9249 for E1, -7.2452 for S4 in OOD), which is impossible for a standard negative log-likelihood of probabilities, since -log p >= 0. This indicates the reported metric is either sign-flipped, computed on logits rather than probabilities, or is some other quantity. This undermines the calibration claims stated in Sections 4.1 and 4.2 and makes the NLL comparisons untrustworthy. The authors must correct the definition and recompute all affected numbers. Relatedly, Eq. (1) is not the RegMixup loss as written: it uses only x_i and y_i on both sides, with no mixup interpolation, despite the text describing interpolation with another sample x_j. Please correct the equation to match the actual loss.
  3. [Section 3.2, Table 1] The description of the N-pair loss weight beta is incomplete: the paper states beta is 0 for 20 epochs, then initialized to 1e-3 and 'updated after every epoch until it reached 0.8 at epoch 50', but does not give the update rule and provides no sensitivity analysis for this schedule. This matters because the claimed benefit of the proposed deep metric loss is not clearly supported by the numbers: S3 (with N-pair) achieves 95.09% in-domain accuracy versus 95.06% for S2 (without N-pair), and lower OOD accuracy (38.89% vs. 39.54%). Please provide the beta schedule, justify its values, and report ablations over beta to demonstrate that the N-pair loss provides the claimed improvement.
  4. [Section 4.3, Eq. (5)] The Frechet Distance in Eq. (5) is computed on the model's own embeddings, and the claim that a low value indicates stable and generalizable features assumes the embedding distributions are approximately Gaussian (the formula is exact only for Gaussian distributions). No evidence for Gaussianity is provided, and deep-network embeddings are typically non-Gaussian and multi-modal. Since Section 4.3 uses Frechet Distance to recommend S3 as the 'most balanced model', the recommendation is unsupported without validating the Gaussian assumption or using a non-parametric distribution distance. Please add diagnostics (e.g., Gaussianity tests or visualizations) or replace the Frechet Distance with a distribution distance that does not require this assumption.
minor comments (5)
  1. [Section 2.2] The abstract and Section 2.2 mention that the N-pair loss is integrated in an 'adversarial setting', but no adversarial training procedure is described anywhere in the paper; please clarify or remove this phrase.
  2. [Section 3.2] The sentence 'All training was conducted using the cross-entropy loss criterion in a multi-class classification setting' appears to contradict the two-stage RE/FD description in Section 2.2, which first uses binary cross-entropy and then RegMixup plus N-pair loss; please align these statements.
  3. [Table 1] Some table cells are misformatted, e.g., '318.16.9395.09' in the S3 row; please fix the spacing and formatting so that all numbers are clearly separated.
  4. [Abstract] The abstract says Frechet Distance is used 'in audio source separation', but the paper is about source tracing, not source separation; please correct the terminology.
  5. [Section 3.2] The hyperparameter description states 'learning rate decay of 0.5' but does not specify the decay schedule (e.g., per epoch or per step); please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained and benchmarked against an external baseline.

full rationale

No load-bearing step reduces to its own inputs. The RE/FD training strategy and NSD detector are imported from Xie et al. [12], which is an external source, and the authors' own prior work ([1], [2], [4]) appears only in peripheral citations. The central comparison is a standard empirical evaluation: models trained with different objectives and backbones are scored on held-out MLAAD evaluation and OOD splits, with Wav2Vec2-AASIST as an external baseline. Fréchet Distance in Eq. (5) is computed from each model's embeddings and used as an evaluation readout; it is not a fitted parameter renamed as a prediction, and no equation equates the reported metrics to training objectives by construction. The claim that lower Fréchet Distance correlates with better performance is an empirical observation from Table 1, not a tautology. Reporting concerns such as the unspecified NSD threshold rule, single-run results, and negative NLL values are correctness or reproducibility risks, not circularity, because they do not show any quantity being defined in terms of the target claim.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities or mechanisms; its contribution is a combination of existing building blocks. The central claim rests on a handful of hand-chosen schedules and borrowed assumptions: the beta warm-up for the N-pair loss, an unstated eta, the dev-tuned NSD threshold, and the Gaussian assumption behind Frechet Distance. None of these is validated with repeated runs or sensitivity analysis.

free parameters (3)
  • N-pair loss weight beta = 0 for the first 20 epochs, then ramps from 1e-3 to 0.8 by epoch 50
    Hand-scheduled weighting of the claimed novel loss component. The schedule directly controls how much the N-pair term contributes, so the improvement attributed to the N-pair loss is only as meaningful as this schedule.
  • RegMixup weight eta = not specified anywhere in the text
    Used in Eq. (1), which is itself printed with two identical CE terms, and never assigned a value, so part of the Fake Dispersion stage loss is not reproducible as described.
  • NSD out-of-domain threshold = determined on the development set, value not reported
    The threshold that decides whether a test sample comes from a novel deepfake algorithm is tuned on the development set (Section 2.3) with no reported value or sensitivity analysis; all OOD accuracy and EER numbers depend on it.
assumptions (3)
  • domain assumption Wav2Vec2-XLSR representations contain enough information to separate 24 generator classes and to generalize to unseen generators.
    Invoked in Section 2.1, where XLSR is chosen as the feature front-end with end-to-end fine-tuning; the whole system inherits whatever bias this pretrained model has.
  • domain assumption Embedding distributions are approximately Gaussian, so the Frechet Distance comparison in Eq. (5) is meaningful.
    Eq. (5) compares in-domain and OOD embeddings by fitted means and covariances; the validity of ranking models by Frechet Distance depends on this untested Gaussian assumption.
  • domain assumption The Real Emphasis and Fake Dispersion two-stage scheme from Xie et al. 2024 transfers to the MLAAD 24-class setup.
    The training recipe is adopted wholesale from the cited baseline; if RE and FD benefits are specific to the ADD2023-T3 setting, the claimed gains partly reflect the borrowed framework rather than the new components.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion." pith.science (2026). https://pith.science/paper/ZAVT3CG7

@misc{pith2026250602085,
  author       = {Pith},
  title        = {Pith review of: Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZAVT3CG7}},
  note         = {Machine review of arXiv:2506.02085}
}
read the original abstract

Audio deepfakes are acquiring an unprecedented level of realism with advanced AI. While current research focuses on discerning real speech from spoofed speech, tracing the source system is equally crucial. This work proposes a novel audio source tracing system combining deep metric multi-class N-pair loss with Real Emphasis and Fake Dispersion framework, a Conformer classification network, and ensemble score-embedding fusion. The N-pair loss improves discriminative ability, while Real Emphasis and Fake Dispersion enhance robustness by focusing on differentiating real and fake speech patterns. The Conformer network captures both global and local dependencies in the audio signal, crucial for source tracing. The proposed ensemble score-embedding fusion shows an optimal trade-off between in-domain and out-of-domain source tracing scenarios. We evaluate our method using Frechet Distance and standard metrics, demonstrating superior performance in source tracing over the baseline system.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 34 canonical work pages

  1. [1]

    Technology has made our lives much easier on one hand, and much harder to believe what we hear and see

    Introduction ’Don’t believe everything you hear’, holds true in today’s highly digitized world. Technology has made our lives much easier on one hand, and much harder to believe what we hear and see. Highly realistic content generated by sophisticated text- to-speech (TTS) [1] and voice conversion systems [2] has led to the growth of deepfake audios, one ...

    work page 2019

  2. [2]

    We begin by providing a brief overview of the classification networks, the feature ex- traction method, and the Deep Neural Network (DNN) mod- els

    Proposed approach In this section, we present the proposed framework for audio source tracing and OOD detection. We begin by providing a brief overview of the classification networks, the feature ex- traction method, and the Deep Neural Network (DNN) mod- els. Next, we elaborate on the training strategies incorporated in our approach, including the novel ...

    work page

  3. [3]

    Dataset This work utilizes the MLAAD dataset [30] for the audio source tracing task

    Experimentation 3.1. Dataset This work utilizes the MLAAD dataset [30] for the audio source tracing task. MLAAD is a multilingual dataset originally de- signed for audio anti-spoofing. It is intended to assess the out- of-domain generalization of anti-spoofing systems across both new languages and unseen TTS models. MLAAD includes378 hours of synthetic sp...

    work page 2019

  4. [4]

    We adopted the evaluation mea- sures stated in Section 3.3

    Results and analysis Table 1 presents the performance of the proposed work in both in-domain and OOD scenarios. We adopted the evaluation mea- sures stated in Section 3.3. In addition to the proposed ap- proach, we experimented with ensemble fusion systems, de- noted byE. For the ensemble fusion, we concatenated the re- spective embeddings from two system...

    work page arXiv

  5. [5]

    The proposed methods showed significant improvements over the baseline in both in-domain and OOD scenarios

    Discussion The evaluation results highlight the system’s performance across different conditions. The proposed methods showed significant improvements over the baseline in both in-domain and OOD scenarios. Among the models proposed, the Con- former model outperformed state-space models like MAMBA and HYDRA. Ablation experiments with three training strate-...

    work page

  6. [6]

    Our experiments demon- strated that while the best-performing models (S2, S3) exhibit strong in-domain performance, challenges persist in generaliz- ing to out-of-domain (OOD) data

    Conclusion In this paper, we presented a two-stage training approach, utiliz- ing Real Emphasis and Fake Dispersion strategies, coupled with the novel multi-class N-pair loss, ensuring robust generaliza- tion across different deepfake sources. Our experiments demon- strated that while the best-performing models (S2, S3) exhibit strong in-domain performanc...

    work page

  7. [7]

    Zhu et.al

    classified spoofing systems by vocoder type. Zhu et.al. [8] attempted multi-attribute classification, including the vocoder, speaker representation, and conversion model. However, distin- guishing sources of synthetic audio remains challenging. Recent advancements in Audio source tracing methods pri- marily focus on in-distribution classification [9]. How...

    arXiv 2023

  8. [8]

    PFFS-21-47), by the Swiss National Science Foundation through the project PASS: Pathological Speech Synthesis (grant agreement no

    Acknowledgements This work was partially supported by the Innosuisse through the flagship project IICT: Inclusive Information and Communi- cation Technologies (grant agreement no. PFFS-21-47), by the Swiss National Science Foundation through the project PASS: Pathological Speech Synthesis (grant agreement no. 219726), and by the Estonian Centre of Excelle...

    work page

Show all 43 references
  1. [9]

    kNN retrieval for simple and effective zero-shot multi-speaker text-to-speech,

    K. E. Hajal, A. Kulkarni, E. Hermann, and M. Magimai Doss, “kNN retrieval for simple and effective zero-shot multi-speaker text-to-speech,” inNAACL, L. Chiruzzo, A. Ritter, and L. Wang, Eds., 2025. [Online]. Available: https://aclanthology.org/2025. naacl-short.65/

    2025

  2. [10]

    Un- supervised rhythm and voice conversion of dysarthric to healthy speech for asr,

    K. E. Hajal, A. Kulkarni, E. Hermann, and M. M. Doss, “Un- supervised rhythm and voice conversion of dysarthric to healthy speech for asr,” inWorkshop on Speech Pathology Analysis and DEtection (SPADE), IEEE, 2025

    2025

  3. [11]

    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. H. Kinnunen, and K. A. Lee, “Asvspoof 2019: Future horizons in spoofed and fake audio detection,” inproc. of Interspeech, 2019

    2019

  4. [12]

    Exploring generalization to unseen au- dio data for spoofing: insights from ssl models,

    A. Kulkarni, H. M. Tran, A. Kulkarni, S. Dowerah, D. Lo- live, and M. M. Doss, “Exploring generalization to unseen au- dio data for spoofing: insights from ssl models,” inThe Auto- matic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024

    2024

  5. [13]

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

    X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. W. D. Evans, M. Sahidullah, V . Vestman, T. H. Kinnunen, K. A. LEE, L. Juvela, P. Alku, Y .-H. Peng, H.-T. Hwang, Y . Tsao, H.-M. Wang, S. L. Maguer, M. Becker, and Z. Ling, “Asvspoof 2019: A large-scale public data...

    2019

  6. [14]

    Synthetic speech detection through short-term and long-term prediction traces,

    C. Borrelli, P. Bestagini, F. Antonacci, A. Sarti, and S. Tubaro, “Synthetic speech detection through short-term and long-term prediction traces,”proc. of EURASIP Journal on Information Se- curity, 2021

    2021

  7. [15]

    Distinguish- ing neural speech synthesis models through fingerprints in speech waveforms,

    C. Y . Zhang, J. Yi, J. Tao, C. Wang, and X. Yan, “Distinguish- ing neural speech synthesis models through fingerprints in speech waveforms,”Proceedings of the 23rd Chinese National Confer- ence on Computational Linguistics, 2024

    2024

  8. [16]

    Source tracing: Detecting voice spoofing,

    T. Zhu, X. Wang, X. Qin, and M. Li, “Source tracing: Detecting voice spoofing,”Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, 2022

    2022

  9. [17]

    Source trac- ing of audio deepfake systems,

    N. Klein, T. Chen, H. Tak, R. Casal, and E. Khoury, “Source trac- ing of audio deepfake systems,”proc. of Interspeech, 2024

    2024

  10. [18]

    Add 2023: the second audio deepfake detection challenge,

    J. Yi, J. Tao, R. Fu, X. Yan, C. Wang, T. Wang, C. Y . Zhang, X. Zhang, Y . Zhao, Y . Ren, and L. Xu, “Add 2023: the second audio deepfake detection challenge,” inproc. of DADA@IJCAI, 2023

    2023

  11. [19]

    Neural codec source tracing: Toward comprehensive attribution in open-set condition,

    Y . Xie, X. Wang, Z. Wang, R. Fu, Z. Wen, S. Cao, L. Ma, C. Li, H. Cheng, and L. Ye, “Neural codec source tracing: Toward comprehensive attribution in open-set condition,” vol. abs/2501.06514, 2025. [Online]. Available: https://arxiv.org/abs/2501.06514

    2025 arXiv

  12. [20]

    Generalized source tracing: Detecting novel audio deep- fake algorithm with real emphasis and fake dispersion strategy,

    Y . Xie, R. Fu, Z. Wen, Z. Wang, X. Wang, H. Cheng, L. Ye, and J. Tao, “Generalized source tracing: Detecting novel audio deep- fake algorithm with real emphasis and fake dispersion strategy,” inproc. of Interspeech, 2024

    2024

  13. [21]

    Detecting domain shift in multiple instance learning for digital pathology using fr ´echet domain distance,

    M. PoceviVCi ¯ut˙e, G. Eilertsen, S. Garvin, and C. Lundstrom, “Detecting domain shift in multiple instance learning for digital pathology using fr ´echet domain distance,” inInternational Con- ference on Medical Image Computing and Computer-Assisted In- tervention, 2024

    2024

  14. [22]

    Fr ´echet video motion distance: A metric for evaluating motion consis- tency in videos,

    J. Liu, Y . Qu, Q. Yan, X. Zeng, L. Wang, and R. Liao, “Fr ´echet video motion distance: A metric for evaluating motion consis- tency in videos,”ArXiv, vol. abs/2407.16124, 2024

    2024 arXiv

  15. [23]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,” in proc. of Interspeech, 2020

    2020

  16. [24]

    Improved deep metric learning with multi-class n-pair loss objective,

    K. Sohn, “Improved deep metric learning with multi-class n-pair loss objective,” inproc. of Neural Information Processing Sys- tems, 2016

    2016

  17. [25]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst Conference on Language Modeling, 2024. [Online]. Available: https://openreview.net/ forum?id=tEYskw1VY2

    2024

  18. [26]

    Hydra: Bidirectional state space models through generalized matrix mixers,

    S. Hwang, A. Lahoti, R. Puduppully, T. Dao, and A. Gu, “Hydra: Bidirectional state space models through generalized matrix mixers,” inproc. of Neural Information Processing Systems,

  19. [27]

    Principles of risk minimization for learning the- ory,

    V . N. Vapnik, “Principles of risk minimization for learning the- ory,” inproc. of Neural Information Processing Systems, 1991

    1991

  20. [28]

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

    A. Baevski, H. Zhou, A. rahman Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”proc. of Neural Information Processing Systems, 2020

    2020

  21. [29]

    Temporal-channel modeling in multi-head self- attention for synthetic speech detection,

    D.-T. Truong, R. Tao, T. Nguyen, H.-T. Luong, K. A. Lee, and E. S. Chng, “Temporal-channel modeling in multi-head self- attention for synthetic speech detection,” inproc. of Interspeech, 2024

    2024

  22. [30]

    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,”Proceedings of the 32nd ACM International Conference on Multimedia, 2024

    2024

  23. [31]

    Xlsr-mamba: A dual-column bidirec- tional state space model for spoofing attack detection,

    Y . Xiao and R. K. Das, “Xlsr-mamba: A dual-column bidirec- tional state space model for spoofing attack detection,”ArXiv, vol. abs/2411.10027, 2024

    2024 arXiv

  24. [32]

    Mfa-conformer: Multi-scale feature aggregation con- former for automatic speaker verification,

    Y . Zhang, Z. Lv, H. Wu, S. Zhang, P. Hu, Z. Wu, H. yi Lee, and H. Meng, “Mfa-conformer: Multi-scale feature aggregation con- former for automatic speaker verification,” inproc. of Interspeech, 2022

    2022

  25. [33]

    Attention is all you need,

    A. Vaswani, N. M. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inproc. of Neural Information Processing Systems, 2017

    2017

  26. [34]

    One-class learning towards syn- thetic voice spoofing detection,

    Y . Zhang, F. Jiang, and Z. Duan, “One-class learning towards syn- thetic voice spoofing detection,”IEEE Signal Processing Letters, 2020

    2020

  27. [35]

    Vicinal risk minimization,

    O. Chapelle, J. Weston, L. Bottou, and V . Vapnik, “Vicinal risk minimization,” inproc. of Neural Information Processing Sys- tems, T. Leen, T. Dietterich, and V . Tresp, Eds., 2000

    2000

  28. [37]

    Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness,

    F. Pinto, H. Yang, S. N. Lim, P. H. S. Torr, and P. K. Dokania, “Using mixup as a regularizer can surprisingly improve accuracy & out-of-distribution robustness,” inproc. of Neural Information Processing Systems, 2022

    2022

  29. [38]

    Deep metric learning: A survey,

    M. Kaya and H. S ¸akir Bilge, “Deep metric learning: A survey,” proc. of Symmetry, vol. 11, p. 1066, 2019

    2019

  30. [39]

    Mlaad: The multi- language audio anti-spoofing dataset,

    N. M. M ¨uller, P. Kawa, W. H. Choong, E. Casanova, E. G ¨olge, T. M¨uller, P. Syga, P. Sperl, and K. B¨ottinger, “Mlaad: The multi- language audio anti-spoofing dataset,”International Joint Confer- ence on Neural Networks (IJCNN), 2024

    2024

  31. [40]

    Using mlaad for source tracing of audio deepfakes,

    N. M ¨uller, “Using mlaad for source tracing of audio deepfakes,” https://deepfake-total.com/sourcetracing, Fraunhofer AISEC, 11 2024

    2024

  32. [41]

    Musan: A music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: A music, speech, and noise corpus,”ArXiv, vol. abs/1510.08484, 2015

    2015 arXiv

  33. [42]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,”proc. of ICASSP, 2017

    2017

  34. [43]

    Frechet mu- sic distance: A metric for generative symbolic music evaluation,

    J. Retkowski, J. Stepniak, and M. Modrzejewski, “Frechet mu- sic distance: A metric for generative symbolic music evaluation,” ArXiv, vol. abs/2412.07948, 2024

    2024 arXiv

  35. [2024]

    Available: https://openreview.net/forum?id= preo49P1VY

    [Online]. Available: https://openreview.net/forum?id= preo49P1VY

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.