pith. sign in

arxiv: 2606.10223 · v1 · pith:T3SPHXUHnew · submitted 2026-06-08 · 💻 cs.SD · cs.AI· cs.CV

Dual-Branch Gated Fusion for Open-Set Audio Deepfake Source Tracing

Awais Khan , Kutub Uddin , Khalid Malik This is my paper

Pith reviewed 2026-06-27 14:43 UTC · model grok-4.3

classification 💻 cs.SD cs.AIcs.CV
keywords audio deepfakesource tracingopen-set recognitiongated fusionXLSR-53CORES descriptordeepfake attributiondistribution shift
0
0 comments X

The pith

An input-conditioned gate fuses XLSR-53 and CORES features to trace deepfake audio sources even when the synthesizer is unseen.

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

The paper establishes that closed-set attribution models fail on new synthesizers by producing overconfident errors, and that a dual-branch architecture can fix this by letting one branch handle in-domain cases while the other handles distribution shifts. It shows that XLSR-53 stays strong inside the training distribution while the 66-dimensional CORES descriptor, built from cepstral, oscillatory, rhythmic, energy, and spectral cues, holds up better outside it. Because direct concatenation fails from representational imbalance, the work introduces an input-dependent gate trained together with cross-entropy, an energy-margin loss, and a diversity term so that the system weights the two branches according to the input. On the MLAAD benchmark this yields 97.6 percent ID accuracy, 4.9 percent EERc, and an 83.5 percent relative drop in FPR95 against the prior Interspeech baseline.

Core claim

The central claim is that an input-conditioned gate, trained jointly with cross-entropy, energy margin loss, and a gate diversity term, adaptively weights the XLSR-53 branch (strong in-domain) against the CORES branch (stable out-of-domain) and thereby resolves representational imbalance that otherwise prevents reliable open-set source tracing of synthetic speech.

What carries the argument

The input-conditioned gate that learns to weight the XLSR-53 and CORES branches on a per-input basis under the combined loss.

If this is right

  • XLSR-53 remains discriminative inside the training distribution while CORES generalizes more stably under shift.
  • Naive concatenation of the two descriptors fails because of SSL representational imbalance.
  • Joint training with cross-entropy, energy margin loss, and gate diversity produces an 83.5 percent relative FPR95 reduction on MLAAD.
  • The same gated system reaches 97.6 percent ID accuracy and 4.9 percent EERc on the benchmark.
  • The framework directly addresses overconfident predictions on unseen synthesizers that closed-set models produce.

Where Pith is reading between the lines

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

  • The same gating idea could be tested on other audio tasks that mix self-supervised embeddings with compact hand-crafted descriptors under distribution shift.
  • If the gate learns to detect when one branch is unreliable, the approach might generalize to multi-modal open-set problems such as video or image forgery attribution.
  • Replacing CORES with alternative low-dimensional descriptors that span similar acoustic dimensions would provide a direct test of whether the 66-dimensional construction is necessary.

Load-bearing premise

That the input-conditioned gate trained with the three losses will reliably balance the two branches without introducing new failure modes on synthesizers never seen during training.

What would settle it

A new collection of synthetic utterances from synthesizers absent from MLAAD on which the gated system shows either lower ID accuracy or higher EERc or FPR95 than the non-gated baseline.

Figures

Figures reproduced from arXiv: 2606.10223 by Awais Khan, Khalid Malik, Kutub Uddin.

Figure 1
Figure 1. Figure 1: Proposed dual-branch gated fusion framework for open-set audio deepfake source tracing. The gating network adaptively balances contextual SSL representations and low-level CORES to improve attribution robustness under unseen generative systems [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation on MLAAD Eval. Single-branch and naive fusion baselines all fail OOD rejection (FPR95 ≥ 34%). Adap￾tive gating resolves the ID/OOD conflict, achieving 87% rela￾tive FPR95 vs. naive concatenation at no ID accuracy cost. On OOD Eval, we obtain 94.3% OOD accuracy and 7.6% EER, surpassing all systems, including the OOD-focused XLSR-HYDRA (S5), by 49.5 pp in OOD accuracy and 47.5 pp in EER, while also … view at source ↗
read the original abstract

Attributing a synthetic utterance to its originating system remains an open challenge: closed-set models fail to reject unseen synthesizers and produce overconfident predictions. To address this, we propose a dual-branch gated fusion framework that pairs XLSR-53 with CORES, a 66-dimensional descriptor that, unlike prior Linear Filter Bank (LFB)-only work, spans cepstral, oscillatory, rhythmic, energy, and spectral dimensions to capture complementary synthesis artifacts. Our analysis shows XLSR-53 remains discriminative in-domain (ID) while CORES generalizes stably under distribution shift (OOD), yet their naive concatenation fails due to SSL representational imbalance. To resolve this, an input-conditioned gate adaptively weights each branch under joint training with cross-entropy, an energy margin loss for ID/OOD separation, and a gate diversity term. On the MLAAD benchmark, our system achieves 97.6\% ID accuracy, 4.9\% EERc, and an 83.5\% relative FPR95 reduction over the Interspeech 2025 baseline.

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 / 0 minor

Summary. The paper proposes a dual-branch gated fusion framework for open-set audio deepfake source tracing that combines XLSR-53 with a new 66-dimensional CORES descriptor spanning cepstral, oscillatory, rhythmic, energy, and spectral features. It argues that XLSR-53 is strong in-domain while CORES generalizes under shift, but naive concatenation fails due to representational imbalance; an input-conditioned gate trained jointly with cross-entropy, energy margin loss, and a gate diversity term adaptively fuses the branches. On the MLAAD benchmark the system reports 97.6% ID accuracy, 4.9% EERc, and an 83.5% relative FPR95 reduction versus the Interspeech 2025 baseline.

Significance. If the headline metrics prove reproducible with proper controls, the work would demonstrate a concrete way to exploit complementary ID/OOD strengths of SSL and hand-crafted descriptors while mitigating imbalance via gating, potentially improving robustness of attribution systems to unseen synthesizers.

major comments (2)
  1. [Abstract] Abstract (framework description): the central claim that the input-conditioned gate, trained with the joint objective, reliably resolves branch imbalance without introducing new failure modes on unseen synthesizers lacks any reported ablation, sensitivity analysis, or theoretical argument showing that the diversity term prevents gate collapse or over-weighting of the less-generalizable branch under distribution shift; this is load-bearing for the reported OOD gains.
  2. [Abstract] Abstract: the performance numbers (97.6% ID accuracy, 4.9% EERc, 83.5% relative FPR95 reduction) are presented without any description of experimental protocol, train/test splits, number of runs, statistical significance tests, or comparison details against the Interspeech 2025 baseline, preventing evaluation of whether the gains are attributable to the proposed gate.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract. We will revise the manuscript to address the concerns about supporting evidence for the gated fusion claims and to include more experimental context in the abstract. Our responses to the major comments are below.

read point-by-point responses

  1. Referee: [Abstract] Abstract (framework description): the central claim that the input-conditioned gate, trained with the joint objective, reliably resolves branch imbalance without introducing new failure modes on unseen synthesizers lacks any reported ablation, sensitivity analysis, or theoretical argument showing that the diversity term prevents gate collapse or over-weighting of the less-generalizable branch under distribution shift; this is load-bearing for the reported OOD gains.

    Authors: We acknowledge that the abstract does not explicitly detail ablations on the diversity term. The full manuscript provides supporting evidence through comparisons of gated fusion versus naive concatenation and individual branches (Section 4.2, Table 2), showing improved OOD metrics. However, to directly address the load-bearing claim regarding prevention of gate collapse under shift, we will add an ablation study on the diversity loss coefficient and gate weight histograms on OOD data in the revised manuscript. This will include sensitivity analysis to demonstrate that the term mitigates over-weighting of the XLSR-53 branch. revision: yes

  2. Referee: [Abstract] Abstract: the performance numbers (97.6% ID accuracy, 4.9% EERc, 83.5% relative FPR95 reduction) are presented without any description of experimental protocol, train/test splits, number of runs, statistical significance tests, or comparison details against the Interspeech 2025 baseline, preventing evaluation of whether the gains are attributable to the proposed gate.

    Authors: The full manuscript details the experimental protocol in Section 3, including MLAAD train/test splits, the Interspeech 2025 baseline re-implementation, and that results are averaged over 5 random seeds. The relative FPR95 reduction is computed as (baseline_FPR95 - our_FPR95) / baseline_FPR95. To improve the abstract, we will add a brief clause summarizing the evaluation setup and note that gains are statistically significant (p < 0.05 via paired t-test). This revision will make the abstract self-contained while preserving length constraints. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes an empirical dual-branch gated fusion architecture pairing XLSR-53 and CORES descriptors, trained jointly under cross-entropy, energy margin, and gate diversity losses, then evaluated on the external MLAAD benchmark against an Interspeech 2025 baseline. No equations, fitted parameters, or self-citations are presented that reduce the reported ID accuracy, EERc, or FPR95 figures to the training inputs by construction; the performance numbers remain independent empirical outcomes rather than tautological renamings or self-referential predictions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review supplies insufficient detail to enumerate free parameters, axioms, or invented entities beyond the high-level description of CORES and the gate; full text would be required for an exhaustive ledger.

invented entities (1)
  • CORES descriptor no independent evidence
    purpose: 66-dimensional feature vector spanning cepstral, oscillatory, rhythmic, energy and spectral dimensions to capture synthesis artifacts
    Introduced in the abstract as a complement to XLSR-53; no independent evidence supplied.

pith-pipeline@v0.9.1-grok · 5722 in / 1347 out tokens · 22057 ms · 2026-06-27T14:43:57.141667+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

44 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    Introduction Hearing is no longer believing.This warning, once reserved for gossip and hearsay, has taken on an unsettling new dimen- sion in the age of generative AI. In January 2024, a phone call impersonating the voice of US President Biden discouraged thousands of New Hampshire voters from attending the primary election, triggering a federal investiga...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    No single feature family captures these reli- ably across both seen and unseen systems

    Proposed Method Our approach is based on a central observation: synthesis arti- facts span multiple representational levels, affecting fine spec- tral cues, harmonic and temporal dynamics, and higher-level phonetic structure. No single feature family captures these reli- ably across both seen and unseen systems. We therefore com- bine two complementary fr...

  3. [3]

    Experiments and Results 3.1. Experimental Setup We evaluate on the MLAAD 2 source tracing protocol [28], which comprises 83 TTS systems across 26 languages parti- tioned into training, development, and evaluation splits (Ta- ble 2). The evaluation set contains 43 entirely unseen syn- thesizers across 26 languages. To enrich OOD coverage, we augment both t...

    2025

  4. [4]

    Ablations confirm that neither SSL nor CORES alone, nor naive concatenation, resolves the ID/OOD trade-off; adaptive gating with energy margin training is essential

    Discussion and Future Work Our results show that exploiting feature complementarity, rather than model scale, enables competitive open-set source tracing. Ablations confirm that neither SSL nor CORES alone, nor naive concatenation, resolves the ID/OOD trade-off; adaptive gating with energy margin training is essential. The consistent gate shift toward COR...

  5. [5]

    (2024, Jan.) Fake biden robocall tells vot- ers to skip new hampshire primary election

    BBC News. (2024, Jan.) Fake biden robocall tells vot- ers to skip new hampshire primary election. Accessed: 2026-03-01. [Online]. Available: https://www.bbc.com/news/ world-us-canada-68064247

    2024

  6. [6]

    J. R. McConvey. (2026, Jan.) Deepfake voice fraud dupes swiss businessman into transferring millions. Accessed: 2026-03-01. [Online]. Available: https://www.biometricupdate.com/202601/ deepfake-voice-fraud-dupes-swiss-businessman

    2026

  7. [7]

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

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

    1989

  8. [8]

    Sheild: A secure and highly enhanced integrated learning for robust deep- fake detection against adversarial attacks,

    K. Uddin, A. Khan, M. U. Farooq, and K. M. Malik, “Sheild: A secure and highly enhanced integrated learning for robust deep- fake detection against adversarial attacks,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 1502–1511

    2025

  9. [9]

    Battling voice spoofing: a review, comparative analysis, and generalizability evaluation of state-of-the-art voice spoofing counter measures,

    A. Khan, K. M. Malik, J. Ryan, and M. Saravanan, “Battling voice spoofing: a review, comparative analysis, and generalizability evaluation of state-of-the-art voice spoofing counter measures,” Artificial Intelligence Review, vol. 56, no. Suppl 1, pp. 513–566, 2023

    2023

  10. [10]

    Advbench: A comprehensive benchmark of adversarial attacks on deepfake detectors in real-world con- sumer applications,

    K. Uddin, M. U. Farooq, A. Khan, M. S. Saeed, I. U. Haq, N. Tas- nim, and K. M. Malik, “Advbench: A comprehensive benchmark of adversarial attacks on deepfake detectors in real-world con- sumer applications,”Authorea Preprints, 2025

    2025

  11. [11]

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

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,”arXiv preprint arXiv:2202.12233, 2022

    work page arXiv 2022

  12. [12]

    Adversarial attacks on audio deepfake detection: A benchmark and compara- tive study,

    K. Uddin, M. U. Farooq, A. Khan, and K. M. Malik, “Adversarial attacks on audio deepfake detection: A benchmark and compara- tive study,”arXiv preprint arXiv:2509.07132, 2025

    work page arXiv 2025

  13. [13]

    Where are we in audio deep- fake detection? a systematic analysis over generative and detec- tion models,

    X. Li, P.-Y . Chen, and W. Wei, “Where are we in audio deep- fake detection? a systematic analysis over generative and detec- tion models,”ACM Transactions on Internet Technology, vol. 25, no. 3, pp. 1–19, 2025

    2025

  14. [14]

    Transfer- able adversarial attacks on audio deepfake detection,

    M. U. Farooq, A. Khan, K. Uddin, and K. M. Malik, “Transfer- able adversarial attacks on audio deepfake detection,” inProceed- ings of the Winter Conference on Applications of Computer Vi- sion, 2025, pp. 1640–1649

    2025

  15. [15]

    Frame-to-utterance con- vergence: A spectra-temporal approach for unified spoofing de- tection,

    A. Khan, K. M. Malik, and S. Nawaz, “Frame-to-utterance con- vergence: A spectra-temporal approach for unified spoofing de- tection,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 10 761–10 765

    2024

  16. [16]

    Interpretable all-type audio deepfake detec- tion with audio llms via frequency-time reinforcement learning,

    Y . Xie, X. Guo, J. Zhou, T. Wang, J. Liu, R. Fu, X. Wang, H. Cheng, and L. Ye, “Interpretable all-type audio deepfake detec- tion with audio llms via frequency-time reinforcement learning,” arXiv preprint arXiv:2601.02983, 2026

    work page arXiv 2026

  17. [17]

    Deepfake algorithm recognition system with augmented data for add 2023 challenge

    X.-M. Zeng, J.-T. Zhang, K. Li, Z.-L. Liu, W.-L. Xie, and Y . Song, “Deepfake algorithm recognition system with augmented data for add 2023 challenge.” inDADA@ IJCAI, 2023, pp. 31–36

    2023

  18. [18]

    Multilingual Source Tracing of Speech Deepfakes: A First Benchmark,

    X. Xuan, Y . Xiao, R. K. Das, and T. Kinnunen, “Multilingual source tracing of speech deepfakes: A first benchmark,”arXiv preprint arXiv:2508.04143, 2025

    work page arXiv 2025

  19. [19]

    Trace: Training- free partial audio deepfake detection via embedding trajectory analysis of speech foundation models,

    A. khan, M. U. Farooq, K. Uddin, and K. Malik, “Trace: Training- free partial audio deepfake detection via embedding trajectory analysis of speech foundation models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) Workshops, June 2026, pp. 7405–7414

    2026

  20. [20]

    Synthetic speech source tracing using metric learning,

    D. Koutsianos, S. Zacharopoulos, Y . Panagakis, and T. Stafylakis, “Synthetic speech source tracing using metric learning,”arXiv preprint arXiv:2506.02590, 2025

    work page arXiv 2025

  21. [21]

    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,” arXiv preprint arXiv:2406.03240, 2024

    work page arXiv 2024

  22. [22]

    The npu-aslp system for deepfake algorithm recognition in add 2023 challenge

    Z. Wang, Q. Wang, J. Yao, and L. Xie, “The npu-aslp system for deepfake algorithm recognition in add 2023 challenge.” in DADA@ IJCAI, 2023, pp. 64–69

    2023

  23. [23]

    Deepfake algorithm recog- nition through multi-model fusion based on manifold measure

    Y . Tian, Y . Chen, Y . Tang, and B. Fu, “Deepfake algorithm recog- nition through multi-model fusion based on manifold measure.” inDADA@ IJCAI, 2023, pp. 76–81

    2023

  24. [24]

    Investigating prosodic signatures via speech pre- trained models for audio deepfake source attribution,

    O. C. Phukan, D. Singh, S. R. Behera, A. B. Buduru, and R. Sharma, “Investigating prosodic signatures via speech pre- trained models for audio deepfake source attribution,” inFind- ings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 4206–4214

    2025

  25. [25]

    Detect- ing unknown speech spoofing algorithms with nearest neighbors

    J. Lu, Y . Zhang, Z. Li, Z. Shang, W. Wang, and P. Zhang, “Detect- ing unknown speech spoofing algorithms with nearest neighbors.” inDADA@ IJCAI, 2023, pp. 89–94

    2023

  26. [26]

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

    C. Zhang, J. Yi, J. Tao, C. Wang, and X. Yan, “Distinguish- ing neural speech synthesis models through fingerprints in speech waveforms,” inProceedings of the 23rd Chinese National Confer- ence on Computational Linguistics (Volume 1: Main Conference), 2024, pp. 1160–1171

    2024

  27. [27]

    Source tracing of audio deepfake systems,

    N. Klein, T. Chen, H. Tak, R. Casal, and E. Khoury, “Source tracing of audio deepfake systems,” inInterspeech

  28. [28]

    1100–1104

    ISCA, Sept 2024, p. 1100–1104. [Online]. Available: http://dx.doi.org/10.21437/Interspeech.2024-1283

    work page doi:10.21437/interspeech.2024-1283 2024

  29. [29]

    Vib-based real pre-emphasis audio deepfake source tracing,

    T.-P. Doan, K. Hong, and S. Jung, “Vib-based real pre-emphasis audio deepfake source tracing,” inProc. Interspeech 2025, 2025, pp. 1568–1572

    2025

  30. [30]

    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 . Renet al., “Add 2023: the second audio deepfake detection challenge,”arXiv preprint arXiv:2305.13774, 2023

    work page arXiv 2023

  31. [31]

    Attacker attribu- tion of audio deepfakes,

    N. M. M ¨uller, F. Dieckmann, and J. Williams, “Attacker attribu- tion of audio deepfakes,”arXiv preprint arXiv:2203.15563, 2022

    work page arXiv 2022

  32. [32]

    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,”arXiv preprint arXiv:2501.06514, 2025

    work page arXiv 2025

  33. [33]

    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

  34. [34]

    Un- veiling audio deepfake origins: A deep metric learning and con- former network approach with ensemble fusion,

    A. Kulkarni, S. Dowerah, T. Alumae, and M. M. Doss, “Un- veiling audio deepfake origins: A deep metric learning and con- former network approach with ensemble fusion,”arXiv preprint arXiv:2506.02085, 2025

    work page arXiv 2025

  35. [35]

    Open-Set Source Tracing of Audio Deepfake Systems,

    N. Klein, H. Tak, and E. Khoury, “Open-Set Source Tracing of Audio Deepfake Systems,” inInterspeech 2025, 2025, pp. 1578– 1582

    2025

  36. [36]

    arXiv preprint arXiv:1904.05862 , year=

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,”arXiv preprint arXiv:1904.05862, 2019

    work page arXiv 1904

  37. [37]

    Xls-r: Self- supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pinoet al., “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

    work page arXiv 2021

  38. [38]

    Securing voice biometrics: One- shot learning approach for audio deepfake detection,

    A. Khan and K. M. Malik, “Securing voice biometrics: One- shot learning approach for audio deepfake detection,” in2023 IEEE international workshop on information forensics and secu- rity (WIFS). IEEE, 2023, pp. 1–6

    2023

  39. [39]

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

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

    2017

  40. [40]

    Audio source identi- fication using delta-delta mfcc features,

    A. Ali, V . Pankajakshan, and S. Sharma, “Audio source identi- fication using delta-delta mfcc features,” in2025 IEEE Interna- tional Conference on Advanced Visual and Signal-Based Systems (AVSS). IEEE, 2025, pp. 1–6

    2025

  41. [41]

    Spotnet: A spoofing-aware trans- former network for effective synthetic speech detection,

    A. Khan and K. M. Malik, “Spotnet: A spoofing-aware trans- former network for effective synthetic speech detection,” inPro- ceedings of the 2nd ACM International Workshop on Multimedia AI against Disinformation, 2023, pp. 10–18

    2023

  42. [42]

    Energy-based out-of- distribution detection,

    W. Liu, X. Wang, J. Owens, and Y . Li, “Energy-based out-of- distribution detection,” inNeurIPS, 2020

    2020

  43. [43]

    Audio deepfake source tracing,

    P. Kawa, “Audio deepfake source tracing,” https://github.com/ piotrkawa/audio-deepfake-source-tracing, 2024, accessed: 2026- 03-02

    2024

  44. [44]

    MUSAN: A Music, Speech, and Noise Corpus

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484v1

    work page internal anchor Pith review Pith/arXiv arXiv 2015