Component-Level Ensemble Fusion for Speech and Environmental Sound Deepfake Detection

Andr\'e Runewicz; Karla Sch\"afer; Martin Steinebach

REVIEW 3 major objections 5 minor 28 references

Component-Level Ensemble Fusion for Speech and Environmental Sound Deepfake Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper reports that a component-wise margin-space ensemble of four pre-trained anti-spoofing models, each with three binary heads and lightweight bias calibration, achieves 0.7828 macro-F1 on the ESDD2 hidden test set, ranking 5th of 31

desk verdict A solid, honest challenge-system paper whose hidden test result is credible and whose marginal calibration claim needs more evidence. read the letter →

arxiv 2607.16369 v1 pith:NPUEXBYP submitted 2026-07-17 cs.SD

André Runewicz , Karla Schäfer , Martin Steinebach This is my paper

classification cs.SD

keywords audiodeepfakedetectioncomponent-levelspoofingenvironmentalsoundensemblefusionmargin-spacepretrainedanti-spoofingmodelsRawBoostaugmentationESDD2challenge

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 for composite audio deepfakes — where speech and environmental sound can be spoofed independently — reusing four pre-trained speech anti-spoofing models as three-headed component detectors and fusing their score margins per component is an effective strategy. On the ESDD2 challenge, the resulting system scores 0.7715 macro-F1 on the evaluation set and 0.7828 on the hidden test set, ranking 5th of 31 and beating the official baseline by about 0.15. The key methodological claim is that component-wise fusion, which lets each model contribute mainly to the head it is best at, outperforms uniform fusion of the same models, and that lightweight head- and class-bias calibration yields further gains. The authors disclose that validation-set performance was not a reliable selector, and they relied on preparation-phase leaderboard feedback for checkpoint and fusion choices.

What carries the argument

The machinery is a three-binary-head factorization on each backbone, producing logit margins d_o, d_s, d_e for original/mixed, speech authenticity, and environment authenticity. Head margins are averaged with component-specific weights (Table III) — e.g., the original head relies mostly on the RawBoost SLS model, the speech head on a DF-Arena run, and the environment head on a balanced four-model mix — then passed through a sigmoid and combined via the factorized probability equations to yield five class scores. Two extra calibration steps adjust decision thresholds: head biases b = (−0.20, −0.30, −0.20) and class biases c = (−0.12, 0, −0.15, −0.15, 0.10).

What would settle it

Fix the eight-checkpoint recipe and all hyperparameters, select checkpoints using validation macro-F1 only without any leaderboard feedback, and score on the same hidden test set; a material drop below 0.7828 macro-F1 would show the result depends on evaluation-fit selection. A second check: bootstrap the 27,603 test clips to estimate a confidence interval for the macro-F1 — a wide interval would weaken the ranking claim.

Watch

Extended reading notes

Core claim

The central discovery is that the five-class component-level deepfake detection problem can be factorized into three binary decisions — original vs mixed, bona fide vs spoofed speech, bona fide vs spoofed environment — and that pre-trained speech anti-spoofing models fine-tuned with three binary heads produce complementary strengths across these components. Combining the head margins with per-head weighted averages (margin-space fusion) and calibrating with a constant head bias plus a class-bias vector gives the best configuration: 0.7715 eval macro-F1 and 0.7828 test macro-F1. This outperforms both the individual models and conservative uniform ensembles on the same backbone set, and beats

Load-bearing premise

The load-bearing premise is that the fusion weights and calibration biases, chosen using preparation-phase leaderboard feedback (Sections III-B and III-H), transfer to the hidden test set; if those choices are overfit to the evaluation split, the reported 0.7828 test macro-F1 and rank-5 result would not be reproducible.

Editorial extensions

If this is right

Component-level factorization with three binary heads transfers pre-trained speech anti-spoofing expertise to environmental sound spoofing without additional training data.
Models show complementary strengths across components, so heterogeneous ensembles are more robust than selecting a single best model by validation score.
RawBoost augmentation helps only some backbones and heads; retaining both augmented and non-augmented checkpoints avoids performance degradation.
Head-bias and class-bias calibration improves macro-F1 by correcting decision thresholds with minimal compute and no retraining.
The final system outperforms the official separation-based baseline mainly through better original/mixed and environment decisions, trading a slightly worse speech EER for higher overall F1.

Reading between the lines

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

Because margin-space fusion outperformed probability-space fusion, calibrating or scaling each model's margins before fusion could yield further gains beyond the fixed weighted average.
The failure of validation-based model selection suggests that adding a domain-adaptation step using unlabeled evaluation-domain audio could reduce the need for leaderboard feedback.
The same three-head factorization could extend to other compositional audio tasks (e.g., music plus lyrics, podcast plus background) and to multimodal deepfake detection, where authenticity of each modality is conditionally independent given a non-original state.
The reported test result is a single submission with no confidence interval; replicating the fixed recipe on a fresh CompSpoofV2-style split would test its stability.

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.

Desk Editor's Note

read the letter

Dear colleague,

Quick take: this ESDD2 submission is a competent, honest system description, and the rank-5 test result (0.7828 macro-F1 vs 0.6327 baseline) is credible. The interesting bit isn't any single component — all the pieces are established — but the three-binary-head factorization for component-level spoofing and the component-wise margin-space fusion with head-specific model weights. That design pattern looks genuinely useful for composite speech/environment deepfake detection.

What the paper does well: the protocol is clearly described, the use of public pre-trained detectors (XLSR-Mamba, DF-Arena, SLS, TCM-ADD) is disclosed, and the authors are transparent that they used preparation-phase leaderboard feedback for checkpoint and fusion selection. The single-model EER table is a nice diagnostic breakdown. The rank-5 result on a hidden test set is an independent external check, not something the authors can manufacture.

The soft spots are real but proportionate. First, the evaluation-set numbers are post-fit by construction: the Table III fusion weights and Section III-H biases were chosen using leaderboard feedback, so the eval table partly reflects those choices. The test set is the honest measure, and that holds up. Second, the calibration gains are small: +0.0037 test macro-F1 for head bias, +0.0025 for class bias. With 27.6k test clips, those deltas are within plausible bootstrap error. So the claim that head/class-bias calibration is the 'best configuration tested' is under-supported — it might be, but the evidence doesn't establish it. Third, no code, seeds, or checkpoints are released, so the selection procedure can't be replayed. That's common in challenge papers, but it does limit verification.

None of this undercuts the central result: the system beats the official baseline by a large margin, and the component-wise fusion strategy outperforms conservative fusion on both eval and test. The paper reads like an honest engineering contribution, not a scientific breakthrough.

Who it's for: anyone working on audio deepfake detection, especially component-level or environment-aware tasks, will find the fusion recipe and the factorization worth studying. It deserves a serious referee; the main revisions I'd ask for are uncertainty quantification (repeated runs or bootstrap CIs), clear labeling of the eval results as post-fit, and ideally code release.

Recommendation: send it to review. It's a competent, useful challenge paper with a credible hidden-test result.

Referee Report

3 major / 5 minor

Summary. The paper reports a system for the ESDD2 challenge's five-class composite audio deepfake detection task, in which speech and environmental sound components may be independently spoofed. The authors fine-tune four publicly available pre-trained anti-spoofing models (XLSR-Mamba, DF-Arena, SLS, TCM-ADD) using three binary heads for original/mixed, speech, and environment decisions, train RawBoost-augmented variants, and combine selected checkpoints via margin-space component-wise score fusion. Two lightweight calibration stages (head biases and class biases) are applied. The submitted configuration achieves 0.7715 macro-F1 on the evaluation set and 0.7828 macro-F1 on the official hidden test set, ranking 5th out of 31 teams and outperforming the official baseline by about 0.15 macro-F1. The paper's central claim is that the component-wise factorized formulation, together with component-wise fusion and bias calibration, is the best configuration tested.

Significance. If the configuration comparison is taken at face value, the paper provides a practical and computationally explicit recipe for reusing open anti-spoofing models in a component-level composite audio task. The test-set result is an externally generated leaderboard number, not a self-reported metric, and the gap over the official baseline is large, so the core approach is credible. The paper also reports a clean ablation structure, gives explicit fusion weights and bias values, and acknowledges the inference cost. The strength of the paper is therefore the external validation of a fairly simple recipe: three binary heads plus margin-space ensemble fusion. The main limitation is that the paper's stronger claim—that the particular fusion recipe with head/class-bias calibration is the best configuration—is not supported by uncertainty quantification and depends on evaluation-leaderboard feedback for model selection. The hidden test result is independent, but the evaluation-set numbers and the ablation ranking are not.

major comments (3)

[Section III-B, III-H, Table V] The evaluation-set rows in Table V are not independent estimates. Section III-B states that candidate checkpoints and fusion recipes were selected using preparation-phase leaderboard feedback, and Section III-H gives head biases and class biases chosen for the component-wise fusion. Thus the evaluation-set macro-F1 values for the calibrated systems are post-selection numbers, not unbiased estimates. The calibration deltas are small: class-bias calibration changes evaluation macro-F1 by +0.0005 (0.7710 to 0.7715) and test macro-F1 by +0.0025 (0.7803 to 0.7828); head-bias calibration changes test macro-F1 by +0.0037. With a single run and no confidence intervals, these differences are comparable to plausible sampling noise. The authors should report bootstrap or permutation confidence intervals on the test-set predictions, state how many leaderboard queries / configurations were explored,
[Section IV-C, Table V] The claim that component-wise fusion with head and class bias is the best configuration tested is the load-bearing configurational claim, but all rows in Table V are single-run realizations. The differences between conservative and component-wise fusion (~0.01 macro-F1) are larger than the calibration deltas but still lack uncertainty quantification. The single test-set realization cannot, by itself, distinguish the calibrated configuration from the head-bias-only configuration. The authors should either provide confidence intervals on the test macro-F1 values, perform a significance test (e.g., block bootstrap over clips or per-class F1 resampling), or rephrase the claim as "the configuration that produced the highest submitted score" rather than "the best configuration tested."
[Reproducibility / artifacts] No code, checkpoints, seeds, or score files are released, so the checkpoint-and-fusion selection procedure described in Section III-B cannot be independently replayed. Because the paper's contribution is essentially a recipe (three binary heads, margin-space fusion, bias calibration), the absence of artifacts weakens the reproducibility of the central claim. Please release at least the fusion/calibration scripts and the exported score files, or state any license/competition restrictions that prevent release, and report the number of random seeds or training runs used for the reported single-model numbers in Table IV.

minor comments (5)

[Section III-G] The statement that "directly averaging margins performed slightly better than probability-space fusion" is not accompanied by quantitative results. Please add a small ablation row or remove the claim.
[Section IV-B] Table IV reports single-model macro-F1 values without any measure of variability. Since these values are used to justify selecting B* as the strongest single model and to discuss RawBoost's inconsistent effect, a repeated-run standard deviation or at least a statement about single-run training would help.
[Section IV-C] The sentence "This suggests that adding an additional model can improve generalization" is inferred from a single cross-over between Conservative-4 and Conservative-5 on evaluation versus test sets. This is too strong for a two-point comparison; please rephrase as an observation.
[Table V caption] The fact that head bias and class bias do not change the EER columns is correctly explained in the text, but a note in the caption would prevent misreading, since the EER columns are identical across the last three rows.
[Abstract / Table VI] The abstract states "5th out of 31 teams," but Table VI lists only the top five participants plus the baseline. Please include the total number of teams in the table caption or text so the rank is self-contained.

Circularity Check

1 steps flagged · score 3.0 of 10

Evaluation-set results are post-fit via leaderboard feedback, but the hidden test-set result is an independent external check; no equation-level circularity.

fitted input called prediction [Section III-B, Section III-H, Tables III and V]
"We therefore export candidate checkpoints after each training epoch and use the preparation-phase leaderboard feedback for checkpoint and fusion selection. ... The strongest configuration is the component-wise fusion with both head-bias and class-bias calibration."

The fusion weights (Table III) and the head/class biases (Section III-H) were selected by using evaluation-set leaderboard feedback on macro-F1. Table V then reports the resulting evaluation-set macro-F1 (0.7715) as evidence that this is the 'strongest configuration.' That evaluation score is a post-fit selection outcome rather than an independent estimate of the configuration's quality. The hidden test-set score (0.7828) was not used in fitting and remains an independent check, so this is a mild, partial circularity rather than a full by-construction identity.

full rationale

The paper's formal derivation is limited to the factorized probability mapping in Eq. (2), which is a decision-level factorization and not a fit-to-target identity. The base detectors are externally pre-trained and fine-tuned only on development data. The one genuinely circular element is the use of the evaluation leaderboard for checkpoint, fusion-weight, and calibration-bias selection, followed by reporting evaluation macro-F1 as evidence for the best configuration. That makes the evaluation-set numbers in Table V optimistic by construction. The test set is held out and provides independent evidence for the headline rank-5 result, and the comparison against the official baseline is external. No load-bearing self-citation chain or uniqueness theorem is present. The lack of confidence intervals affects robustness of the configuration-level claim but is not itself a circularity. Overall, the circularity is low-to-moderate and confined to the post-fit evaluation-set reporting.

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

The system has no new theoretical entities. Its contribution rests on three fitted components (fusion weights, head biases, class biases) plus four domain assumptions. The fitted quantities are responsible for most of the reported improvement over the uncalibrated ensemble, and the factorized independence assumption is load-bearing for the five-class probabilities.

free parameters (5)

Component-wise fusion weights (Table III) = Original: C_RB=0.70, A_RB=0.15, B_RB=0.15; Speech: B*=0.959, B=0.027, B_RB=0.014; Environment: C=0.25, D_RB=0.25, A=0.25
Selected empirically using preparation-phase evaluation leaderboard feedback; no principled derivation is given.
Head biases b_o, b_s, b_e = -0.20, -0.30, -0.20
Tuned on the evaluation set to improve macro-F1 (Section III-H).
Class biases c_0..c_4 = -0.12, 0.00, -0.15, -0.15, 0.10
Tuned on the evaluation set for the final five-class argmax (Section III-H).
Original-head class weights = [0.2 mixed, 0.8 original]
Hand-chosen class weighting in the original/mixed loss (Section III-D).
Loss weights lambda_o, lambda_s, lambda_e = 1.0, 1.0, 1.0
Hand-chosen; not swept (Section III-D).

assumptions (4)

domain assumption Five-class probabilities factorize as P(S,E|O=0) ≈ P(S|O=0)P(E|O=0) in Eq. (2).
Used to build p0-p4 from marginal heads; if speech and environment authenticity are dependent given a mixed clip, class probabilities are misspecified. The paper says this is a decision-level factorization, not an independence claim, but the equations still use the product.
domain assumption Speech-trained SSL anti-spoofing models transfer to environmental-sound spoofing detection.
All four backbones are speech anti-spoofing detectors; the environment head is added without architecture designed for acoustic scenes (Section III-C).
ad hoc to paper Evaluation-leaderboard feedback is a trustworthy guide for final checkpoints, weights, and biases.
Sections III-B and III-H select checkpoints, fusion weights, and biases using the preparation-phase evaluation set; the test score is the only truly independent check, and there is no uncertainty analysis.
domain assumption Raw logit margins from different architectures are numerically comparable for weighted averaging in Eq. (5).
Margins d_h from different backbones are averaged directly without per-model normalization or calibration before fusion.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Component-Level Ensemble Fusion for Speech and Environmental Sound Deepfake Detection." pith.science (2026). https://pith.science/paper/NPUEXBYP

@misc{pith2026260716369,
  author       = {Pith},
  title        = {Pith review of: Component-Level Ensemble Fusion for Speech and Environmental Sound Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NPUEXBYP}},
  note         = {Machine review of arXiv:2607.16369}
}

read the original abstract

This paper describes our submission to the ICME 2026 ESDD2 challenge on environment-aware speech and sound deepfake detection. The task requires five-class classification of audio clips in which speech, environmental sound, both components, or neither component may be spoofed. We propose a component-level ensemble system based on four publicly available pre-trained anti-spoofing models: XLSR-Mamba, DF-Arena, SLS, and TCM-ADD. Each model is fine-tuned on the official CompSpoofV2 development data using three binary heads for original, speech, and environmental sound detection. We further train RawBoost-augmented variants and combine selected checkpoints using margin-space score fusion. A component-wise fusion strategy with lightweight head- and class-bias calibration yields our best configuration, reaching 0.7715 macro-F1 on the evaluation set and 0.7828 macro-F1 on the test set, ranking 5th out of 31 teams in the final ranking phase and substantially outperforming the official baseline.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 linked inside Pith

[1]

Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,”arXiv preprint arXiv:2303.03926, 2023

arXiv 2023
[2]

Esdd2: Environment-aware speech and sound deepfake detection challenge evaluation plan,

Xueping Zhang, Han Yin, Yang Xiao, Lin Zhang, Ting Dang, Ro- han Kumar Das, and Ming Li, “Esdd2: Environment-aware speech and sound deepfake detection challenge evaluation plan,”arXiv preprint arXiv:2601.07303, 2026

arXiv 2026
[3]

Compspoof: A dataset and joint learning framework for component- level audio anti-spoofing countermeasures,

Xueping Zhang, Yechen Wang, Linxi Li, Liwei Jin, and Ming Li, “Compspoof: A dataset and joint learning framework for component- level audio anti-spoofing countermeasures,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 18067–18071

2026
[4]

Esdd2-compspoof-v2: A compos- ite spoofing dataset for speech anti-spoofing,

Xueping Zhang and Ming Li, “Esdd2-compspoof-v2: A compos- ite spoofing dataset for speech anti-spoofing,” https://huggingface. co/datasets/XuepingZhang/ESDD2-CompSpoof-V2, 2025, Accessed: 2026-04-24

2025
[5]

Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,

Andreas Nautsch, Xin Wang, Nicholas Evans, Tomi H. Kinnunen, Ville Vestman, Massimiliano Todisco, H ´ector Delgado, Md Sahidullah, Junichi Yamagishi, and Kong Aik Lee, “Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,”IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 3, no. 2, pp....

2019
[6]

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

Junichi Yamagishi, Xin Wang, Massimiliano Todisco, Md Sahidullah, Jose Patino, Andreas Nautsch, Xuechen Liu, Kong Aik Lee, Tomi Kin- nunen, Nicholas Evans, et al., “Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,”2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 2021

2021
[7]

ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,

Xin Wang, H ´ector Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidul- lah, Tomi H. Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi, “ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,” inThe Automatic Speaker Verification Spoofing Countermeasures Wor...

2024
[8]

Towards end- to-end synthetic speech detection,

Guang Hua, Andrew Beng Jin Teoh, and Haijian Zhang, “Towards end- to-end synthetic speech detection,”IEEE Signal Processing Letters, vol. 28, pp. 1265–1269, 2021

2021

Show all 28 references

[9]

End-to-end anti-spoofing with rawnet2,

Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher, “End-to-end anti-spoofing with rawnet2,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6369–6373

2021
[10]

Does audio deepfake detection generalize?,

Nicolas Michael M ¨uller, Pavel Czempin, Franziska Dieckmann, Adam Froghyar, and Konstantin B ¨ottinger, “Does audio deepfake detection generalize?,” inInterspeech, 2022

2022
[11]

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

Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”Advances in neural information processing systems, vol. 33, pp. 12449–12460, 2020

2020
[12]

Hubert: Self- supervised speech representation learning by masked prediction of hidden units,

Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakho- tia, Ruslan Salakhutdinov, and Abdelrahman Mohamed, “Hubert: Self- supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM Transactions on Audio, Speech, and Language Processing,...

2021
[13]

XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,

Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli, “XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale,” ...

2022
[14]

Automatic speaker verification spoof- ing and deepfake detection using wav2vec 2.0 and data augmentation,

Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee-weon Jung, Junichi Yamagishi, and Nicholas Evans, “Automatic speaker verification spoof- ing and deepfake detection using wav2vec 2.0 and data augmentation,” Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 2022

2022
[15]

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

Yang Xiao and Rohan Kumar Das, “Xlsr-mamba: A dual-column bidirectional state space model for spoofing attack detection,”IEEE Signal Processing Letters, 2025

2025
[16]

Audio deepfake detection with self-supervised xls-r and sls classifier,

Qishan Zhang, Shuangbing Wen, and Tao Hu, “Audio deepfake detection with self-supervised xls-r and sls classifier,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 6765–6773

2024
[17]

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

Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng, “Temporal-channel modeling in multi-head self-attention for synthetic speech detection,” inInterspeech 2024, 2024, pp. 537–541

2024
[18]

Scenefake: An initial dataset and benchmarks for scene fake audio detection,

Jiangyan Yi, Chenglong Wang, Jianhua Tao, Chu Yuan Zhang, Cunhang Fan, Zhengkun Tian, Haoxin Ma, and Ruibo Fu, “Scenefake: An initial dataset and benchmarks for scene fake audio detection,”Pattern Recognition, vol. 152, pp. 110468, 2024

2024
[19]

Envfake: An initial environmental-fake audio dataset for scene-consistency detec- tion,

Hannan Cheng, Kangyue Li, Long Ye, and Jingling Wang, “Envfake: An initial environmental-fake audio dataset for scene-consistency detec- tion,” in2024 IEEE 14th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2024, pp. 81–85

2024
[20]

Detection of deepfake environmental audio,

Hafsa Ouajdi, Oussama Hadder, Modan Tailleur, Mathieu Lagrange, and Laurie M Heller, “Detection of deepfake environmental audio,” in2024 32nd European Signal Processing Conference (EUSIPCO). IEEE, 2024, pp. 196–200

2024
[21]

Fakesound: Deepfake general audio detection,

Zeyu Xie, Baihan Li, Xuenan Xu, Zheng Liang, Kai Yu, and Mengyue Wu, “Fakesound: Deepfake general audio detection,” inProc. Inter- speech 2024, 2024

2024
[22]

Audiocaps: Generating captions for audios in the wild,

Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim, “Audiocaps: Generating captions for audios in the wild,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1...

2019
[23]

Envsdd: Benchmarking envi- ronmental sound deepfake detection,

Han Yin, Yang Xiao, Rohan Kumar Das, Jisheng Bai, Haohe Liu, Wenwu Wang, and Mark D Plumbley, “Envsdd: Benchmarking envi- ronmental sound deepfake detection,” inProc. Interspeech 2025, 2025, pp. 201–205

2025
[24]

Does audio deepfake detection rely on artifacts?,

Tsu-Hsien Shih, Chin-Yuan Yeh, and Ming-Syan Chen, “Does audio deepfake detection rely on artifacts?,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 12446–12450

2024
[25]

Audio deepfake detection under post-processing attack,

Karla Sch ¨afer, Jeong-Eun Choi, and Martin Steinebach, “Audio deepfake detection under post-processing attack,” in2025 33rd European Signal Processing Conference (EUSIPCO). IEEE, 2025, pp. 586–590

2025
[26]

Do compact ssl backbones matter for audio deepfake detection? a controlled study with raptor,

Ajinkya Kulkarni, Sandipana Dowerah, Atharva Kulkarni, Tanel Alum¨ae, and Mathew Magimai Doss, “Do compact ssl backbones matter for audio deepfake detection? a controlled study with raptor,” 2026

2026
[27]

Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,

Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, and Nicholas Evans, “Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” 2022

2022
[28]

Decoupled weight decay regulariza- tion,

Ilya Loshchilov and Frank Hutter, “Decoupled weight decay regulariza- tion,” inInternational Conference on Learning Representations, 2019

2019

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

Ziqiang Zhang, Long Zhou, Chengyi Wang, Sanyuan Chen, Yu Wu, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,”arXiv preprint arXiv:2303.03926, 2023

arXiv 2023

[2] [2]

Esdd2: Environment-aware speech and sound deepfake detection challenge evaluation plan,

Xueping Zhang, Han Yin, Yang Xiao, Lin Zhang, Ting Dang, Ro- han Kumar Das, and Ming Li, “Esdd2: Environment-aware speech and sound deepfake detection challenge evaluation plan,”arXiv preprint arXiv:2601.07303, 2026

arXiv 2026

[3] [3]

Compspoof: A dataset and joint learning framework for component- level audio anti-spoofing countermeasures,

Xueping Zhang, Yechen Wang, Linxi Li, Liwei Jin, and Ming Li, “Compspoof: A dataset and joint learning framework for component- level audio anti-spoofing countermeasures,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 18067–18071

2026

[4] [4]

Esdd2-compspoof-v2: A compos- ite spoofing dataset for speech anti-spoofing,

Xueping Zhang and Ming Li, “Esdd2-compspoof-v2: A compos- ite spoofing dataset for speech anti-spoofing,” https://huggingface. co/datasets/XuepingZhang/ESDD2-CompSpoof-V2, 2025, Accessed: 2026-04-24

2025

[5] [5]

Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,

Andreas Nautsch, Xin Wang, Nicholas Evans, Tomi H. Kinnunen, Ville Vestman, Massimiliano Todisco, H ´ector Delgado, Md Sahidullah, Junichi Yamagishi, and Kong Aik Lee, “Asvspoof 2019: Spoofing countermeasures for the detection of synthesized, converted and replayed speech,”IEEE Transactions on Biometrics, Behavior, and Identity Science, vol. 3, no. 2, pp....

2019

[6] [6]

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

Junichi Yamagishi, Xin Wang, Massimiliano Todisco, Md Sahidullah, Jose Patino, Andreas Nautsch, Xuechen Liu, Kong Aik Lee, Tomi Kin- nunen, Nicholas Evans, et al., “Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection,”2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, 2021

2021

[7] [7]

ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,

Xin Wang, H ´ector Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidul- lah, Tomi H. Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi, “ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale,” inThe Automatic Speaker Verification Spoofing Countermeasures Wor...

2024

[8] [8]

Towards end- to-end synthetic speech detection,

Guang Hua, Andrew Beng Jin Teoh, and Haijian Zhang, “Towards end- to-end synthetic speech detection,”IEEE Signal Processing Letters, vol. 28, pp. 1265–1269, 2021

2021