Pith. sign in

REVIEW 2 major objections 4 minor 42 references

Hidden-Domain Routing for All-Type Audio Deepfake Detection

T0 review · 2 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read All-type audio deepfake detection improves when a system first recovers the hidden audio domain and then applies a domain-specific detector and decision rule, reaching 96.10% Track2 Macro-F1 and first place on the official AT-ADD Track2 lea

desk verdict Sound competition report with an externally verified first-place result, but the paper overstates how much of the gain is due to routing per se rather than to the stronger EAT experts and fusion rules. read the letter →

arxiv 2608.00493 v1 pith:NR6HF267 submitted 2026-08-01 cs.SD

Yifan Gao , Yao Tian , Hongbin Suo , Haonan Lu This is my paper
classification cs.SD
keywords audiodeepfakedetectionhidden-domainconditionaudio-typeroutingall-typedomain-conditionedBEATsEATAT-ADDTrack2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper targets all-type audio deepfake detection, where a clip may be speech, environmental sound, singing, or music but the audio type is hidden at inference. It argues that one generic detector is the wrong tool for this task, because representation geometry and score meaning differ across audio types. The proposed system first predicts the audio type with a BEATs-based router, then runs only the matching branch and applies branch-specific thresholds and fusion rules. On the official blind final evaluation it reports 96.10% Track2 Macro-F1, ranking first and beating the strongest official baseline by 16.63 points. The type-wise scores are 88.07% for speech, 98.18% for sound, 99.07% for singing, and 99.08% for music, with speech remaining the clear bottleneck.

What carries the argument

The load-bearing component is the AudioType-BEATs-6s Router, a four-way audio-type classifier built on a frozen BEATs embedding with a 6-second input window. Its top-1 prediction selects one branch, and each branch interprets scores locally: the Speech-XLSR Expert (XLS-R frontend plus AASIST backend) for speech, and EAT-based SoundMusic-EAT and Singing-EAT experts for non-speech domains. The branch-local decision rules, tuned separately per type on development data, are what make routing matter, since a score reliable for speech is not assumed to carry the same semantics for sound, singing, or music.

What would settle it

Rerun the same pipeline on the official final evaluation with the router disabled, letting a single best expert decide all clips; if the Track2 Macro-F1 gap to 96.10% shrinks or disappears, domain routing is not the cause. If per-sample final labels become available, compute type-wise Macro-F1 separately for correctly and incorrectly routed clips, since the central claim predicts that misrouted clips score far worse.

Watch

Extended reading notes

Core claim

The central claim is that shared binary real/fake labels across speech, sound, singing, and music hide a domain condition: a detector score is not type-independent. The system therefore recovers the audio type first with the AudioType-BEATs-6s Router, then interprets detector outputs within the selected branch using deterministic branch-local rules, including five-crop conjunctive rules for speech and singing and an OR-fake fusion for sound and music. On the official AT-ADD Track2 final evaluation, this routed system achieves 96.10% Track2 Macro-F1 and ranks first, with type-wise Macro-F1 of 88.07%, 98.18%, 99.07%, and 99.08% for speech, sound, singing, and music.

Load-bearing premise

The load-bearing premise is that the per-branch thresholds and decision rules chosen on development subsets to maximize type-wise Macro-F1 will transfer to the blind final evaluation distribution; if that transfer fails, the reported margins, especially the speech branch's 88.07%, could change.

Editorial extensions

If this is right

  • Recovering the hidden audio type before judging real versus fake appears to be the decisive design choice, giving a 16.63-point Track2 Macro-F1 gain over the strongest non-routed official baseline.
  • Sound and music detection reach above 98% type-wise Macro-F1, suggesting these non-speech domains are close to saturated on the current evaluation distribution, while speech at 88.07% is the remaining hard case.
  • Type-wise Macro-F1 is necessary for interpreting all-type audio detection, because the aggregate Track2 score hides the speech bottleneck.
  • A router plus specialized experts is a workable closed-track recipe even with limited per-resource training, relying on frozen self-supervised frontends rather than large new models.

Reading between the lines

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

  • The router is used as a hard gate, so a misrouted clip can never be judged by the right expert; a soft version that mixes branch scores by router probability is a natural extension the paper does not explore.
  • Because the thresholds and fusion rules were selected on development subsets, their transfer to unseen generators or to audio types not in the four-class taxonomy is untested and could be a source of fragility.
  • The fixed 6-second router window matches vocal-domain durations but under-represents longer music and sound contexts; a length-adaptive router could change routing accuracy and downstream scores on long clips.
  • The final-set t-SNE and silhouette values use router-predicted types, not ground truth, so the final-set geometry does not independently verify domain separation; that check would require per-sample labels.
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

2 major / 4 minor

Summary. The paper addresses AT-ADD Track2, where systems output a binary real/fake label for clips from speech, sound, singing, and music without the clip type at inference time. The authors formalize this as binary detection under a hidden audio-domain condition and propose a routed system: a BEATs-based 6-second audio-type router selects one of four branches; speech uses an XLS-R/AASIST expert, while sound, singing, and music use EAT-based experts. Each branch applies per-type thresholds and special aggregation rules (five-crop conjunctive for speech/singing; OR-fake fusion of two EAT experts for sound/music). The paper reports development-set router-family comparisons, representation geometry, and component results, plus an official final evaluation score of 96.10% Track2 Macro-F1, ranking first, with type-wise Macro-F1 values of 88.07%, 98.18%, 99.07%, and 99.08% for speech, sound, singing, and music.

Significance. The official result is externally validated and substantial: 96.10% Track2 Macro-F1 is a 16.63-point gain over FT-XLSR-AASIST, and the type-wise reporting usefully exposes speech as the bottleneck. The paper is transparent about the blind evaluation and its limitations (Section 4.5). If the routing mechanism were causally established, the work would make a useful contribution to domain-conditioned audio deepfake detection. However, the current evidence does not isolate routing from the simultaneous introduction of EAT experts, five-crop aggregation, and OR-fake fusion; the central attribution therefore needs an additional control experiment.

major comments (2)
  1. [Section 3.4 and Table 3] The central claim that 'recovering the hidden audio domain before interpreting detector scores' drives the improvement is not yet supported, because the routed system differs from standalone experts in two coupled ways: it adds the top-1 router and it changes the decision rule (5-crop conjunctive for speech/singing; OR-fake fusion of SoundMusic-EAT and Singing-EAT for sound/music). Progress data show SoundMusic-EAT already reaches 97.70/98.99 on sound/music, close to the routed 98.49/99.02; the +3.50 gain over the best standalone expert could come from the fusion/aggregation rules or the stronger EAT backbone rather than from routing. Please add a no-router control on the development set (e.g., all experts with a global threshold, and an oracle-router variant using ground-truth type), plus a factorized comparison without the fusion rules while keeping the router. Without this, the causal
  2. [Section 3.4 and Section 4.5] All decision thresholds and the OR-fake/5-crop rules are selected on the development subsets to maximize the exact type-wise Macro-F1, and all reported Progress/Eval numbers are single point estimates with no error bars or significance tests. Because final per-sample labels are unavailable, the stability of the speech branch's 88.07% and of the claimed margins cannot be assessed. This is not a circularity problem, but it is a transfer-risk concern. Please report a development-internal split or bootstrap over dev clips to show threshold/fusion-rule stability, and state explicitly which design choices were frozen before any Progress feedback; at minimum, also provide a router confusion matrix on the development set.
minor comments (4)
  1. [Abstract / Section 3.2] The abstract uses 'closed-condition' while the body uses 'closed track'; unify the terminology for consistency.
  2. [Section 4.2 and Figure 2] The silhouette values 0.104/0.084 are low; the claim that samples 'form type-dependent regions' should be calibrated to this magnitude. The final-evaluation panel is colored by router-predicted type and is therefore a self-partition, not independent validation; the text acknowledges this, but the caption could state it explicitly.
  3. [Section 4.4] Since Progress is a subset of Eval, comparing '96.29 on Progress' with '96.10 on Eval' is a consistency check, not a generalization result. The wording 'only 0.19 percentage points lower' should be adjusted accordingly.
  4. [Section 3.2 / Section 4.3] The router uses a 6-second window and 'audio longer than 6 seconds is cropped to the first 6 seconds'; please clarify the handling of clips shorter than 6 seconds (padding, truncation, or rejection), since Table 1 shows mean durations near 5 seconds for speech and singing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline result is an externally scored blind-evaluation result; dev-set threshold fitting and self-cited references do not make the derivation circular.

full rationale

The paper's central result (96.10% Track2 Macro-F1, Rank 1) comes from the official AT-ADD Track2 final evaluation, which is blind at the sample level (Section 4.1: 'The evaluation data are blind at the sample level: participants submit binary predictions, and the organizers return aggregate scores.'). The router, branch experts, and decision rules are all trained and selected using official train/dev labels; the dev-set threshold fitting (Section 3.4: 'Decision thresholds are selected separately for each audio-type branch by maximizing the corresponding type-wise Macro-F1 on the development subset') is standard supervised model selection, not a fitted-input-called-prediction step, because the final-set labels are never used. The paper does not lean on self-citations for load-bearing claims: references to AT-ADD, EAT, XLS-R, AASIST, and related work are to external sources with no author overlap, and no uniqueness theorem or ansatz is imported from prior work by the same authors. The one methodological gap—the absence of a no-router control that keeps the same experts and fusion rules—is a causal-attribution/ablation concern, not circularity: it does not make any equation or score equal to its own input by construction. Section 4.2 also explicitly warns that the final-evaluation t-SNE is colored by router-predicted type and 'it is not an independent validation of audio-type separation,' so the paper does not present a self-confirming analysis as independent evidence. Therefore no circular step is exhibited.

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

The central claim rests on the official benchmark score and on the assumption that development-tuned thresholds transfer. The free parameters are the per-branch decision thresholds and hand-chosen rules; no invented entities are introduced.

free parameters (5)
  • Branch decision thresholds (speech/sound/singing/music) = Not reported; selected to maximize dev type-wise Macro-F1
    Each branch's real/fake cutoff is tuned on the corresponding development subset (Section 3.4), so the final scores depend on these fitted cutoffs.
  • Router input window (6 s) = 6 seconds
    Hand-picked as near the mean duration of speech/singing clips (Section 4.2); affects router accuracy and hence which branch thresholds apply.
  • Temporal crop counts = 5 (speech), 5 (singing), 1 (sound and music)
    Hand-chosen inference aggregation rule; the 5-crop conjunctive rule and single-crop scoring are part of the branch-local decision function (Sections 3.3 and 3.4).
  • OR-fake fusion rule for sound and music = Fake if either expert says fake
    Hand-chosen decision rule (Section 3.3) that biases predictions toward the majority fake class; its threshold interaction affects type-wise F1.
  • Training crop durations and music sampling weight = 4s (speech), 10s (EAT); increased music sampling
    Training recipe choices (Section 4.3) that determine detector behavior and therefore scores.
assumptions (5)
  • domain assumption Audio type is a meaningful latent domain that changes representation geometry and detector-score semantics
    This is the paper's core framing (Section 3.1). It is supported only by the dev geometry (silhouette 0.104, Section 4.2), not by a formal argument.
  • domain assumption The router trained on official dev labels maintains accuracy on the blind evaluation set
    Branch selection depends on router reliability; final-set router accuracy is unknown because per-sample eval labels are hidden (Section 4.5).
  • domain assumption Dev-set-optimal decision thresholds and fusion rules transfer to the evaluation distribution
    Section 3.4 states thresholds are picked on dev subsets; the final score assumes transfer. The paper's Section 4.5 acknowledges final labels are unavailable, so this cannot be validated.
  • standard math Macro-F1 is computed as the unweighted average of per-type real/fake F1 means
    Defined in Section 4.1; this is the evaluation metric from the AT-ADD protocol.
  • domain assumption Augmentation of bona fide speech with TTS-like artifacts produces training examples that generalize to real fake speech
    The Speech-XLSR Expert's augmentation strategy (Section 4.3) assumes artificial artifacts approximate generator outputs; otherwise speech accuracy would not generalize.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hidden-Domain Routing for All-Type Audio Deepfake Detection." pith.science (2026). https://pith.science/paper/NR6HF267

@misc{pith2026260800493,
  author       = {Pith},
  title        = {Pith review of: Hidden-Domain Routing for All-Type Audio Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NR6HF267}},
  note         = {Machine review of arXiv:2608.00493}
}
read the original abstract

All-type audio deepfake detection requires authenticity decisions across speech, environmental sound, singing voice, and music, while the audio type is unavailable at inference time. In AT-ADD Track2, this setting creates a hidden audio-domain condition: the binary real/fake label is shared across domains, but representation structure and detector-score behavior vary with audio type. We present a closed-condition routed system that first recovers the hidden audio domain and then interprets detector scores within the selected branch. The AudioType-BEATs-6s Router estimates audio type from a 6-second window; speech inputs are handled by the Speech-XLSR Expert, while sound, singing, and music rely on EAT-based general-audio experts with branch-local score interpretation. Development-set representation analysis, router-family comparisons, and component results show audio-domain separation and complementary detector strengths across audio types. On the official AT-ADD Track2 final evaluation, the system achieves 96.10% Track2 Macro-F1 and ranks first on the final leaderboard, with type-wise Macro-F1 scores of 88.07%, 98.18%, 99.07%, and 99.08% for speech, sound, singing, and music, respectively. These results support recovering the hidden audio domain before interpreting detector scores in all-type audio deepfake detection.

Figures

Figures reproduced from arXiv: 2608.00493 by the authors.

Figure 1
Figure 1. Routed all-type audio deepfake detection pipeline. The AudioType-BEATs-6s Router predicts the hidden audio type, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Router-domain representation geometry for AT-ADD Track2. Both panels use one t-SNE projection of balanced [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 15 canonical work pages

  1. [1]

    Martin Arjovsky et al . 2019. Invariant Risk Minimization. arXiv:1907.02893 doi:10.48550/arXiv.1907.02893

    arXiv doi:10.48550/arxiv.1907.02893 2019

  2. [2]

    Arun Babu et al. 2022. XLS-R: Self-Supervised Cross-Lingual Speech Representa- tion Learning at Scale. InProceedings of Interspeech 2022. ISCA, Incheon, Republic of Korea, 2278–2282. doi:10.21437/Interspeech.2022-143

    doi:10.21437/interspeech.2022-143 2022

  3. [3]

    Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Represen- tations. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., Red Hook, NY, USA, 12449–12460. arXiv:2006.11477 [cs.CL] doi:10.48550/arXiv.2006.11477

    arXiv doi:10.48550/arxiv.2006.11477 2020

  4. [4]

    Yuxuan Cao et al. 2026. Efficient Audio Transformer and AASIST for Environment Sound Deepfake Detection in the ESDD 2026 Challenge. InProceedings of ICASSP. IEEE, Barcelona, Spain, 21781–21783. doi:10.1109/ICASSP55912.2026.11464115

    arXiv 2026

  5. [5]

    Sanyuan Chen et al. 2022. WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing.IEEE Journal of Selected Topics in Signal Processing 16, 6 (2022), 1505–1518. doi:10.1109/JSTSP.2022.3188113

    arXiv 2022

  6. [6]

    Sanyuan Chen et al. 2023. BEATs: Audio Pre-Training with Acoustic Tokenizers. InProceedings of the International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 202). PMLR, Honolulu, HI, USA, 5178–5193. arXiv:2212.09058 https://proceedings.mlr.press/v202/chen23ag.html

    arXiv 2023

  7. [7]

    Wenxi Chen, Yuzhe Liang, Ziyang Ma, Zhisheng Zheng, and Xie Chen. 2024. EAT: Self-Supervised Pre-Training with Efficient Audio Transformer. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. Interna- tional Joint Conferences on Artificial Intelligence Organization, Jeju, Republic of Korea, 3807–3815. arXiv:2401.03...

    arXiv 2024

  8. [8]

    Yu-An Chung et al . 2021. W2v-BERT: Combining Contrastive Learning and Masked Language Modeling for Self-Supervised Speech Pre-Training. InProceed- ings of ASRU. IEEE, Cartagena, Colombia, 244–250. doi:10.1109/ASRU51503.2021. 9688253

    arXiv 2021

Show all 42 references
  1. [9]

    Lorenzo Comanducci et al . 2024. FakeMusicCaps: A Dataset for Detec- tion and Attribution of Synthetic Music Generated via Text-to-Music Models. arXiv:2409.10684 doi:10.48550/arXiv.2409.10684

    2024 doi arXiv

  2. [10]

    Benjamin Elizalde et al. 2023. CLAP: Learning Audio Concepts from Natural Language Supervision. InProceedings of ICASSP. IEEE, Rhodes Island, Greece, 1–5. doi:10.1109/ICASSP49357.2023.10095889

    2023

  3. [11]

    Yaroslav Ganin et al. 2016. Domain-Adversarial Training of Neural Networks. Journal of Machine Learning Research17, 59 (2016), 1–35. https://www.jmlr.org/ papers/v17/15-239.html

    2016

  4. [12]

    Yuan Gong, Yu-An Chung, and James Glass. 2021. AST: Audio Spectrogram Transformer. InProceedings of Interspeech 2021. ISCA, Brno, Czechia, 571–575. doi:10.21437/Interspeech.2021-698

    2021 doi

  5. [13]

    Ishaan Gulrajani and David Lopez-Paz. 2021. In Search of Lost Domain General- ization. InProceedings of the International Conference on Learning Representations. OpenReview.net, Virtual Event. arXiv:2007.01434 doi:10.48550/arXiv.2007.01434

    2021 doi arXiv

  6. [14]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On Calibration of Modern Neural Networks. InProceedings of the International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 70). PMLR, Sydney, Australia, 1321–1330. arXiv:1706.0459...

    2017 doi arXiv

  7. [15]

    Yuxuan Hao et al. 2025. ALLM4ADD: Unlocking the Capabilities of Audio Large Language Models for Audio Deepfake Detection. InProceedings of the 33rd ACM International Conference on Multimedia(Dublin, Ireland)(MM ’25). Association for Computing Machinery, New York, NY, USA, 1173...

    2025 doi

  8. [16]

    Wei-Ning Hsu et al . 2021. HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units.IEEE/ACM Transactions on Audio, Speech, and Language Processing29 (2021), 3451–3460. doi:10.1109/TASLP. 2021.3122291

    2021

  9. [17]

    Po-Yao Huang et al. 2022. Masked Autoencoders that Listen. InAdvances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., Red Hook, NY, USA, 28708–28720. doi:10.52202/068431-2081

    2022 doi

  10. [18]

    Jacobs, Michael I

    Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton

  11. [19]

    Jordan and Robert A

    Michael I. Jordan and Robert A. Jacobs. 1994. Hierarchical Mixtures of Experts and the EM Algorithm.Neural Computation6, 2 (1994), 181–214. doi:10.1162/ neco.1994.6.2.181

    1994

  12. [20]

    Jee-weon Jung et al. 2019. RawNet: Advanced End-to-End Deep Neural Network Using Raw Waveforms for Text-Independent Speaker Verification. InProceedings of Interspeech 2019. ISCA, Graz, Austria, 1268–1272. doi:10.21437/Interspeech. 2019-1982

    2019 doi

  13. [23]

    Yiming Li et al. 2024. Cross-Domain Audio Deepfake Detection: Dataset and Analysis. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Miami, FL, USA, 4977–4983. doi:10.18653/v1/2024.emnlp-main.286

    2024 doi

  14. [24]

    Lipton, Yu-Xiang Wang, and Alexander J

    Zachary C. Lipton, Yu-Xiang Wang, and Alexander J. Smola. 2018. Detecting and Correcting for Label Shift with Black Box Predictors. InProceedings of the International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80). PMLR, Stockholm, Sweden, 3...

    2018 doi arXiv

  15. [25]

    Nicolas Mueller et al. 2024. Harder or Different? Understanding Generalization of Audio Deepfake Detection. InProceedings of Interspeech 2024. ISCA, Kos Island, Greece, 2705–2709. doi:10.21437/Interspeech.2024-247

    2024 doi

  16. [26]

    Muhammad Mustafa et al. 2025. Multi-level SSL Feature Gating for Audio Deep- fake Detection. InProceedings of the 33rd ACM International Conference on Multi- media(Dublin, Ireland)(MM ’25). Association for Computing Machinery, New York, NY, USA, 11766–11775. doi:10.1145/374602...

    2025

  17. [27]

    Andreas Nautsch et al. 2021. ASVspoof 2019: Spoofing Countermeasures for the Detection of Synthesized, Converted and Replayed Speech.IEEE Transactions on Biometrics, Behavior, and Identity Science3, 2 (2021), 252–265. doi:10.1109/ TBIOM.2021.3059479

    2021

  18. [28]

    Lawrence (Eds.)

    Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D. Lawrence (Eds.). 2009.Dataset Shift in Machine Learning. MIT Press, Cambridge, MA, USA. doi:10.7551/mitpress/9780262170055.001.0001

    2009

  19. [29]

    Noam Shazeer et al. 2017. Outrageously Large Neural Networks: The Sparsely- Gated Mixture-of-Experts Layer. InProceedings of the International Conference on Learning Representations. OpenReview.net, Toulon, France. arXiv:1701.06538 doi:10.48550/arXiv.1701.06538

    2017 doi arXiv

  20. [30]

    Hemlata Tak et al. 2022. Automatic Speaker Verification Spoofing and Deepfake Detection Using wav2vec 2.0 and Data Augmentation. InProceedings of the Odyssey Speaker and Language Recognition Workshop. ISCA, Beijing, China, 112–

    2022

  21. [31]

    Hemlata Tak et al. 2022. RawBoost: A Raw Data Boosting and Augmentation Method Applied to Automatic Speaker Verification Anti-Spoofing. InProceedings of ICASSP. IEEE, Singapore, 6382–6386. doi:10.1109/ICASSP43922.2022.9746213

    2022

  22. [32]

    Yuxiang Xie et al. 2024. Domain Generalization via Aggregation and Separation for Audio Deepfake Detection.IEEE Transactions on Information Forensics and Security19 (2024), 344–358. doi:10.1109/TIFS.2023.3324724

    2024

  23. [33]

    Zhifeng Xie et al. 2026. AT-ADD: All-Type Audio Deepfake Detection Challenge Evaluation Plan. arXiv:2604.08184 [cs.SD] doi:10.48550/arXiv.2604.08184

    2026 doi arXiv

  24. [34]

    Zhifeng Xie et al. 2026. Detect All-Type Deepfake Audio: Wavelet Prompt Tuning for Enhanced Auditory Perception. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 40. AAAI Press, Palo Alto, CA, USA, 35922–35930. doi:10.1609/aaai.v40i42.40907

    2026 doi

  25. [35]

    Junichi Yamagishi et al. 2021. ASVspoof 2021: Accelerating Progress in Spoofed and Deepfake Speech Detection. arXiv:2109.00537 doi:10.48550/arXiv.2109.00537

    2021 doi arXiv

  26. [36]

    Jiangyan Yi et al. 2022. ADD 2022: The First Audio Deep Synthesis Detection Challenge. arXiv:2202.08433 doi:10.48550/arXiv.2202.08433

    2022 doi arXiv

  27. [37]

    Jiangyan Yi et al . 2023. ADD 2023: The Second Audio Deepfake Detection Challenge. arXiv:2305.13774 doi:10.48550/arXiv.2305.13774

    2023 doi arXiv

  28. [38]

    Ziqi Yin et al. 2025. EnvSDD: Benchmarking Environmental Sound Deepfake Detection. InProceedings of Interspeech 2025. ISCA, Rotterdam, The Netherlands, 201–205. doi:10.21437/Interspeech.2025-1143

    2025 doi

  29. [39]

    Ziqi Yin et al . 2025. ESDD 2026: Environmental Sound Deepfake Detection Challenge Evaluation Plan. arXiv:2508.04529 doi:10.48550/arXiv.2508.04529

    2025 doi

  30. [40]

    Yuan Zang et al . 2024. SingFake: Singing Voice Deepfake Detection. arXiv:2309.07525 doi:10.48550/arXiv.2309.07525

    2024 doi

  31. [41]

    Qi Zhang, Haolin Wen, and Bin Hu. 2024. Audio Deepfake Detection with Self- Supervised XLS-R and SLS Classifier. InProceedings of the 32nd ACM International Conference on Multimedia(Melbourne, VIC, Australia)(MM ’24). Association for Computing Machinery, New York, NY, USA, 676...

    2024 doi

  32. [42]

    You Zhang et al . 2024. SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge. arXiv:2408.16132 doi:10.48550/arXiv.2408.16132

    2024 doi

  33. [119]

    doi:10.21437/Odyssey.2022-16

    2022 doi

  34. [1991]

    doi:10.1162/neco.1991.3.1.79

    Adaptive Mixtures of Local Experts.Neural Computation3, 1 (1991), 79–87. doi:10.1162/neco.1991.3.1.79

    1991 doi

Pith tools

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