The DOI matches its Crossref record; 40 of the 40 references it cites were found in a registry or an index; 12 of the 16 sources its claims cite were checked for support.
WavLM model ensemble for audio deepfake detection
Open accessAbstract
Audio deepfake detection has become a pivotal task over the last couple of years, as many recent speech synthesis and voice cloning systems generate highly realistic speech samples, thus enabling their use in malicious activities. In this paper we address the issue of audio deepfake detection as it was set in the ASVspoof5 challenge. First, we benchmark ten types of pretrained representations and show that the self-supervised representations stemming from the wav2vec2 and wavLM families perform best. Of the two, wavLM is better when restricting the pretraining data to LibriSpeech, as required by the challenge rules. To further improve performance, we finetune the wavLM model for the deepfake detection task. We extend the ASVspoof5 dataset with samples from other deepfake detection datasets and apply data augmentation. Our final challenge submission consists of a late fusion combination of four models and achieves an equal error rate of 6.56% and 17.08% on the two evaluation sets.
Overview
CoChat Summary
TL;DR
The paper benchmarks pretrained audio representations for ASVspoof5 audio deepfake detection and finds wavLM-base best among challenge-compliant models, then finetunes it with external data and augmentation. A late-fusion ensemble of four wavLM-base variants achieves 6.56% EER on the progress phase and 17.08% EER on the final evaluation set.
Key findings
- In linear probing on a 27k ASV5 development subset, wavLM-base achieved 9.93% EER, the best among challenge-compliant models; wav2vec2-base reached 13.33% and HuBERT 16.47%. Non-compliant models wavLM-large and wav2vec2-xls-r-2b reached 6.67% and 0.96% EER, respectively.
- Finetuning improved wavLM-base: medium-27k gave 4.16% dev EER; augm-31k gave 0.61% dev and 9.02% progress EER; augm-114k gave 2.97% dev and 7.26% progress EER. Finetuned wav2vec2-base on medium-27k gave 5.85% dev and 11.91% progress EER.
- The final late-fusion ensemble of four wavLM-base variants—pretrained plus finetuned on medium-27k, augm-31k, and augm-114k—achieved 6.56% EER on the progress phase and 17.08% EER on the final evaluation set.
- On the development set, the best two-model combination was augm-31k plus augm-114k, with 0.56% EER; data augmentation was an important component for finetuning.
Methods
- The work addressed the ASVspoof5 open deepfake detection track. Ten pretrained representations from self-supervised, speaker embedding, and learnable-frontend classes were benchmarked as frozen feature extractors with logistic regression over average-pooled representations, using two random 27k subsets from ASV5 train/dev with an 8:1 spoof-to-bonafide ratio.
- wavLM-base and wav2vec2-base were finetuned with binary cross-entropy, Adam at learning rate 3·10⁻⁵, linear scheduler with 0.1 warmup, 5 epochs, batch size 8 or 16, on a single Tesla V100 16GB GPU. Training data included ASV5 plus ASV19, ASV21, FoR, and ITW, with white-noise and reverberation augmentation applied to half of bonafide samples; dataset variants were medium-27k, augm-31k, and augm-114k.
- The final submission used late fusion: logistic regression over model output probabilities, with weights learned on the medium-27k train split, evaluated on development, progress phase, and final evaluation sets.
Limitations
- Stated: limited computational resources prevented finetuning on the complete datasets; not all progress phase evaluation results could be presented because the phase was closed by submission time.
- Evident: the final evaluation EER of 17.08% is much higher than the progress phase EER of 6.56%, indicating a substantial generalization gap to the final evaluation set.
- Evident: the ASVspoof5 open-condition restrictions excluded top-performing models such as wav2vec2-xls-r-2b and wavLM-large, so the best challenge-compliant model was weaker than those non-compliant toplines.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
3 Valid · 1 Not valid · 0 Not checked · 16 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“However, the limitation within the ASV5 challenge's open track discarded the majority of the top performing readily available pretrained models.”
What the sources say
- Disputedcites [20] [20] X. Wang et al. , 'ASVspoof 5: Crowdsourced data, deepfakes and adversarial attacks at scale,' in ASVspoof 2024 Workshop (submitted) , 2024. - sourceExperimental — accuracy not yet measured
Our copy describes the open condition as the track that ALLOWS pre-trained speech foundation models (Introduction, §1; §2 data protocol: 'participants in the open condition were permitted to use external training data, but only under the condition that there is no overlap with the challenge database'), not one that discarded them. §6.1 (Track 1) states the contrary outcome directly: 'most of the top-performing submissions use features extracted using pre-trained, self-supervised learning (SSL) models, e.g., wav2vec 2.0 (base version)', and open-condition minDCFs/EERs were lower (better) than closed. The paper also shows the no-overlap rule was accommodated by removing LibriSpeech-overlapping speakers from the MLS evaluation data so that pre-trained models remained usable (§ near Conclusions), and nowhere does it say the majority of top-performing readily available pretrained models were ruled out. (LLM, conf=0.72)
Checked by a reasoning model — may be wrong.
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“As shown above, large SSL models have shown very good performance over the deepfake detection task.”
What the sources say
- Not checked — no full text of the sourcecites [1] [1] X. Wang and J. Yamagishi, 'Investigating self-supervised front ends for speech spoofing countermeasures,' in Proc. Odyssey: The Speaker and Language Recognition Workshop , 2022. - source
- Not checked yetcites [2] [2] J. M. Mart´ ın-Do˜ nas and A. ´ Alvarez, 'The Vicomtech audio deepfake detection system based on wav2vec2 for the 2022 ADD challenge,' in Proc. ICASSP , 2022, pp. 9241-9245. -
- Supportedcites [3] [3] H. Tak, M. Todisco, X. Wang, J. Jung, J. Yamagishi, and N. W. D. Evans, 'Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation,' in Proc. Odyssey: The Speaker and Language Recognition Workshop , 2022. - source
The Related work section identifies the large wav2vec 2.0 XLS-R (0.3B) model used in the study. The Conclusions report up to an 88% relative reduction in equal error rate on a domain-mismatched deepfake detection task, supporting the claim of very good performance, while cautioning that comparisons with most prior results are not fair because training data differed. (LLM, conf=0.91)
- Supportedcites [4] [4] Y. Yang, H. Qin, H. Zhou, C. Wang, T. Guo, K. Han, and Y. Wang, 'A robust audio deepfake detection system via multi-view feature,' in Proc. ICASSP , 2024, pp. 13 13113 135. - source
The introduction (p. 1) says self-supervised audio features are beneficial for audio deepfake detection. The Experiments section (p. 3) reports strong results for the large pretrained Wav2Vec2 XLS-R model and excellent In-the-Wild performance for HuBERT and WavLM features. (LLM, conf=0.92)
- Supportedcites [5] [5] Y. Guo, H. Huang, X. Chen, H. Zhao, and Y. Wang, 'Audio deepfake detection with self-supervised wavLM and multi-fusion attentive classifier,' in Proc. ICASSP , 2024, pp. 12 702-12 706. - source
The Introduction reports that self-supervised speech models performed quite well as spoofing-detection front ends. Section 4.1 reports that the paper’s WavLM-Large system achieved the best performance on the ASVspoof 2021 deepfake evaluation set among the compared systems. (LLM, conf=0.90)
- Not checked — no full text of the sourcecites [6] [6] H. Wu, J. Zhang, Z. Zhang, W. Zhao, B. Gu, and W. Guo, 'Robust spoof speech detection based on multi-scale feature aggregation and dynamic convolution,' in Proc. ICASSP , 2024, pp. 10 156-10 160. - source
- Supportedcites [7] [7] S. Saha, M. Sahidullah, and S. Das, 'Exploring green AI for audio deepfake detection,' CoRR , vol. abs/2403.14290, 2024. - source
In “D. Towards Green AI solution” (page unavailable), the paper reports a 0.90% equal error rate for its wav2vec 2.0-based deepfake detector and says other studies achieved even lower error rates using larger models. This supports the broad claim that large SSL-based models have performed very well at the task. (LLM, conf=0.88)
- Supportedcites [8] [8] O. Pascu, A. Stan, D. Oneata, E. Oneata, and H. Cucu, 'Towards generalisable and calibrated synthetic speech detection with self-supervised representations,' in Proc. Interspeech , 2024. - source
Section 4.1 reports that a large pretrained self-supervised representation achieved substantially better average audio deepfake-detection performance than RawNet2 (8.8% versus 30.9% equal error rate). The claim is a fair broad summary of that result, although the tested approach pairs a frozen representation with a classifier. (LLM, conf=0.92)
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Among the three wavLM-base fi netuning variants, the ones using augmented data perform best: the agum-31k set gives the lowest error on the development set (0.61% EER), while the augm-114k set gives the lowest error on the progress phase evaluation set (7.26% EER).”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“On the development set, we observe improvements by using model combinations, with the best combination of two consisting of the models finetuned on the largest datasets: augm-31k and augm-114k (row 4).”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Of the two, wavLM is better when restricting the pretraining data to LibriSpeech, as required by the challenge rules.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
8 more claims
Create a free account to see all 20 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 1
- It may state two separate claims as one. 4
- The quote only partly states this claim. 4
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The cited source says what the claim attributes to it.
In the citing paper
“In particular, the proposed Fused SSL feature + Improved NeXt-TDNN approach attains an EER of 7.23%, which outperforms even its closest competitor, WavLM (Combei et al., 2024 [45]).”
From Robust DeepFake Audio Detection via an Improved NeXt-TDNN with Multi-Fused Self-Supervised Learning Features (2025) · cites this paper as [45]
Citations
Cites
“Librispeech: an ASR corpus based on public domain audio books,”
Panayotov, V., Chen, G., Povey, D., & Khudanpur, S. (2015) · 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2015.7178964 · Cited by 6,072 · cited 2×
“As topline, we also include results for two models trained on LibriLight [21] or Multilingual LibriSpeech [24], which were consequently not allowed in the challenge: wavLM-large and wav2vec2-xls-r-2b .”
MethodResultAudio Set: An ontology and human-labeled dataset for audio events
Gemmeke, J. F., Ellis, D. P. W., Freedman, D., Jansen, A., Lawrence, W., Moore, R. C., Plakal, M., & Ritter, M. (2017) · 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2017.7952261 · Cited by 3,134 · cited 1×
“This model was trained on Audioset-2M [28] which also includes non-speech audio.”
MethodHuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units
Hsu, W.-N., Bolte, B., Tsai, Y.-H. H., Lakhotia, K., Salakhutdinov, R., & Mohamed, A. (2021) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2021.3122291 · Cited by 2,983 · cited 2×
“• HuBERT [11] is a Hidden-Unit BERT approach for self-supervised speech representation learning.”
Methodwav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations
Baevski, A., Zhou, H., Mohamed, A., & Auli, M. (2020) · arXiv · doi:10.48550/arxiv.2006.11477 · Cited by 2,369 · cited 3×
“Methods such as wav2vec [10], HuBERT [11] or wavLM [12] achieve this desideratum by reconstructing masked parts of the input audio.”
BackgroundMethodWavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing
Chen, S., Wang, C., Chen, Z., Wu, Y., Liu, S., Chen, Z., Li, J., Kanda, N., Yoshioka, T., Xiao, X., et al. (2022) · IEEE Journal of Selected Topics in Signal Processing · doi:10.1109/jstsp.2022.3188113 · Cited by 1,926 · cited 2×
“Methods such as wav2vec [10], HuBERT [11] or wavLM [12] achieve this desideratum by reconstructing masked parts of the input audio.”
BackgroundMethodDesplanques, B., Thienpondt, J., & Demuynck, K. (2020) · Interspeech 2020 · doi:10.21437/interspeech.2020-2650 · Cited by 1,568 · cited 1×
“We selected the ECAPA-TDNN [30] and TitaNet [31] models.”
MethodPerceptual Signal Analysis and Eigenvalue-Weighted Metrics for Musical Audio Quality Assessment
Majumder, S. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-27827-2_35 · Cited by 845 · cited 1×
“In the open condition, there was a single limitation pertaining to the use of models or datasets which included samples from the LibriLight [21], Multilingual LibriSpeech [22] or MUSAN [23] datasets.”
MethodSUPERB: Speech Processing Universal PERformance Benchmark
Yang, S.-W., Chi, P.-H., Chuang, Y.-S., Lai, C.-I. J., Lakhotia, K., Lin, Y. Y., Liu, A. T., Shi, J., Chang, X., Lin, G.-T., et al. (2021) · Interspeech 2021 · doi:10.21437/interspeech.2021-1775 · Cited by 681 · cited 1×
“The resulting representations can be successfully employed by multiple downstream tasks (e.g. speech recognition, keyword spotting, speaker identification) with limited data [13].”
BackgroundUnsupervised Cross-Lingual Representation Learning for Speech Recognition
Conneau, A., Baevski, A., Collobert, R., Mohamed, A., & Auli, M. (2021) · Interspeech 2021 · doi:10.21437/interspeech.2021-329 · Cited by 637 · cited 1×
“The wav2vec family of models [10,16,17] remains the most popular option for deepfake detection [1–3,7,8,18,19].”
BackgroundXLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale
Babu, A., Wang, C., Tjandra, A., Lakhotia, K., Xu, Q., Goyal, N., Singh, K., von Platen, P., Saraf, Y., Pino, J., et al. (2022) · Interspeech 2022 · doi:10.21437/interspeech.2022-143 · Cited by 586 · cited 2×
“The XLS-R 2B variant of wav2vec [17] is a large-scale model for cross-lingual speech representation learning based on wav2vec 2.0.”
BackgroundMethodInfluential
Cited by
Tahaoglu, G. (2025) · Applied Sciences · doi:10.3390/app15179685 · Cited by 11
Ali, H., Subramani, S., Adupa, N. S., Bollinani, L., El-Loh, S., & Malik, H. (2025) · IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU) · doi:10.1109/asru65441.2025.11434712 · Cited by 6
“Recent works [31], [32]”
Post-training for Deepfake Speech Detection
Ge, W., Wang, X., Liu, X., & Yamagishi, J. (2025) · IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU) · doi:10.1109/asru65441.2025.11434709 · Cited by 36
“Therefore, the representations extracted from pre-trained SSL models are normally further optimized using domain specific data during the fine-tuning phase every time for each deepfake dataset [1], [3], [9], [10].”
A Comprehensive Approach to Deepfake Audio Detection: Using Feature Fusion and Deep Learning
Momu, S. A., Siddiqui, R. R., Shanto, S. S., & Ahmed, Z. (2024) · International Conference on Computer and Information Technology (ICCIT) · doi:10.1109/iccit64611.2024.11022599 · Cited by 2
“In this study [9] the author utilized the wavLM model to improve audio deepfake detection.”
Ameer, N. H. A., Al-Taee, M. M., Hussain, A.-S. T., Hussian, A., & Haider, A. (2025) · IEEE International Conference on Computing and Machine Intelligence (ICMI) · doi:10.1109/icmi65310.2025.11141228 · Cited by 2
Metadata
- Authors
- Combei, D., Stan, A., Oneata, D., & Cucu, H.
- Venue
- The Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024)
- Publisher
- ISCA
- Published
- 2024-08-31
- Pages
- 170-175
- OpenAlex
- W4403956249
- Cited by
- 15
- References
- 37
- Citations by year
- 2026: 7 · 2025: 7 · 2024: 1
- FWCI
- 4.43
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- public-domain
- Locations
- Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof) (publishedVersion)
- arXiv (Cornell University) (submittedVersion, public-domain)
- arXiv (Cornell University)
- Funders
- Unitatea Executiva pentru Finantarea Invatamantului Superior, a Cercetarii, Dezvoltarii si Inovarii
- Colegiul Consultativ pentru Cercetare-Dezvoltare şi Inovare
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
- Physical Sciences › Computer Science › Signal Processing › Music and Audio Processing
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Keywords
- audio deepfake detection, WavLM, ensemble learning, audio classification, audio forensics
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: WavLM model ensemble for audio deepfake detection. https://app.cochat.ai/papers/10.21437/asvspoof.2024-25
- The paper
- Combei, D., Stan, A., Oneata, D., & Cucu, H. (2024). WavLM model ensemble for audio deepfake detection. The Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024). https://doi.org/10.21437/asvspoof.2024-25
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.