The DOI matches its Crossref record; 32 of the 32 references it cites were found in a registry or an index; 18 of the 20 sources its claims cite were checked for support.
Audio Deepfake Detection With Self-Supervised Wavlm And Multi-Fusion Attentive Classifier
Open accessAbstract
With the rapid development of speech synthesis and voice conversion technologies, Audio Deepfake has become a serious threat to the Automatic Speaker Verification (ASV) system. Numerous countermeasures are proposed to detect this type of attack. In this paper, we report our efforts to combine the self-supervised WavLM model and Multi-Fusion Attentive classifier for audio deepfake detection. Our method exploits the WavLM model to extract features that are more conducive to spoofing detection for the first time. Then, we propose a novel Multi-Fusion Attentive (MFA) classifier based on the Attentive Statistics Pooling (ASP) layer. The MFA captures the complementary information of audio features at both time and layer levels. Experiments demonstrate that our methods achieve state-of-the-art results on the ASVspoof 2021 DF set and provide competitive results on the ASVspoof 2019 and 2021 LA set.
Overview
CoChat Summary
## TL;DR
The paper proposes an audio deepfake detector that uses WavLM as a self-supervised front-end and a new Multi-Fusion Attentive (MFA) classifier to combine time- and layer-level information. It reports state-of-the-art results on ASVspoof 2021 DF and competitive results on ASVspoof 2019/2021 LA.
## Key findings
- On the ASVspoof 2021 DF evaluation set, the proposed method with WavLM-Large and MFA achieves a pooled EER of 2.56%, which the paper describes as the lowest reported EER on that set.
- On the ASVspoof 2021 LA evaluation set, the method achieves a pooled EER of 5.08%.
- On the ASVspoof 2019 LA evaluation set, it achieves a pooled min t-DCF of 0.0126 and EER of 0.42%.
- In ablation results, using WavLM-Large without fine-tuning still gives 5.79% EER, and removing MFA hurts performance: 6.01% EER with GAP and 2.92% EER with TN + FCs.
## Methods
- The system feeds raw waveforms into WavLM to extract multi-layer hidden representations, then sends those representations to an ASP-based Multi-Fusion Attentive (MFA) classifier for final prediction.
- Experiments use ASVspoof 2019 LA train/dev for training and evaluate on ASVspoof 2019 LA, ASVspoof 2021 LA, and ASVspoof 2021 DF evaluation sets.
- The main analysis reports pooled EER and min t-DCF, along with an ablation study over front-end choice, fine-tuning, and classifier variants.
## Limitations
- The paper does not report any explicit limitations section; this is not stated in the paper.
- From the text, the approach is evaluated only on ASVspoof 2019/2021 partitions, so broader generalization beyond those datasets is not demonstrated in the paper.
- The paper also notes that some results are only “competitive” on LA, indicating performance is not uniformly best across all test sets; this is evident from the reported results.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
15 Valid · 0 Not valid · 2 Not checked · 19 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Attributed to the masked speech prediction and denoising training, WavLM has the potential to learn nonASR features, including complex acoustic environments and speaker-related information and had demonstrated improved performance on certain non-ASR tasks.”
What the sources say
- Supportedcites [9] source
The Introduction (p. 2) reports improved results for non-ASR tasks including speaker verification, speech separation, and speaker diarization. The Conclusion (p. 11) explicitly says WavLM performs well on both ASR and non-ASR tasks. (LLM, conf=0.94)
boostedstates 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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“we draw inspiration from recent advancements in speech self-supervised model, WavLM [9] and introduce the usage of WavLM as a front-end feature extractor for the first time.”
What the sources say
- Supportedcites [9] source
The Introduction (p. 1) presents WavLM as a self-supervised speech model, and the Speech Separation section (p. 10) describes using its representations as inputs to downstream models. The citation identifies the model that inspired the citing authors; their claim to be the first to use it for audio deepfake detection is their own, not a claim attributed to this source. (LLM, conf=0.91)
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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Attributed to the masked speech prediction and denoising training, WavLM has the potential to learn nonASR features, including complex acoustic environments and speaker-related information and had demonstrated improved performance on certain non-ASR tasks.”
What the sources say
- Supportedcites [9] source
The Introduction (p. 1) says masked speech prediction and denoising help WavLM learn non-ASR information, including speaker identity and the ability to handle noisy or overlapping speech. The Introduction (p. 2) reports improved results on speaker verification, speech separation, and diarization tasks. (LLM, conf=0.93)
boostedstates a possible 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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Previous work [5, 12, 13] has demonstrated that the intermediate feature of self-supervised model contains certain discriminative information.”
What the sources say
- Supportedcites [5] source
Section 2.2 (p. 2) says first or intermediate layers can provide more discriminative information for tasks such as speaker verification and emotion recognition. The Conclusion (p. 4) says this paper exploits representations from different transformer layers to detect spoofed speech, supporting the cited claim. (LLM, conf=0.93)
- Supportedcites [12] source
Section 5.3 (p. 4) reports that hidden-layer representations contain speaker information and that lower-layer information is more discriminative for automatic speaker verification than higher-layer information. This supports the claim that intermediate features of self-supervised models contain discriminative information. (LLM, conf=0.94)
- Supportedcites [13] source
In Results and discussion (page not specified in our copy), the paper reports that combining intermediate wav2vec 2.0 layers improved speech-emotion recognition over using only encoder outputs. It also reports that the trained model gave middle layers larger weights. This supports the claim that intermediate features of a self-supervised model contain discriminative information for this task. (LLM, conf=0.91)
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.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“However, features from current speech self-supervised models are inadequate for multi-speaker tasks [9], including audio deepfake detection.”
What the sources say
- Partly supportedcites [9] sourceExperimental — accuracy not yet measured
The Introduction (p. 1) says existing pre-trained models are unsatisfactory for multi-speaker tasks, specifically speaker diarization and speech separation. It does not identify audio deepfake detection as a multi-speaker task or assess model features for that application. (LLM, conf=0.88)
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.
- Partly supported
12 more claims · 1 with their cited sources checked
Create a free account to see all 36 claims and every source verdict.
- It may state two separate claims as one. 12
- Probably not a substantial claim. 8
- The quote only partly states this claim. 3
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 source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“WavLM and HuBERT representations have also been employed [1,4,5,8], but their results seem to trail those of wav2vec.”
From WavLM model ensemble for audio deepfake detection (2024) · cites this paper as [5]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Guo et al. [129] proposed a multifusion attentive (MFA) classifier to extract and integrate outputs from different Transformer layers and time steps of WavLM. Their results highlighted the potential of SSL models in audio deepfake detection, particularly in challenging acoustic environments.”
From Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead (2025) · cites this paper as [129]
The cited source says what the claim attributes to it.
In the citing paper
“As shown above, large SSL models have shown very good performance over the deepfake detection task.”
From WavLM model ensemble for audio deepfake detection (2024) · cites this paper as [5]
Citations
Cites
wav2vec 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
Desplanques, B., Thienpondt, J., & Demuynck, K. (2020) · Interspeech · doi:10.48550/arxiv.2005.07143 · Cited by 2,295
WavLM: 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
Desplanques, B., Thienpondt, J., & Demuynck, K. (2020) · Interspeech 2020 · doi:10.21437/interspeech.2020-2650 · Cited by 1,568
ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection
Todisco, M., Wang, X., Vestman, V., Sahidullah, M., Delgado, H., Nautsch, A., Yamagishi, J., Evans, N., Kinnunen, T., & Lee, K. A. (2019) · arXiv · doi:10.48550/arxiv.1904.05441 · Cited by 688
ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection
Todisco, M., Wang, X., Vestman, V., Sahidullah, M., Delgado, H., Nautsch, A., Yamagishi, J., Evans, N., Kinnunen, T. H., & Lee, K. A. (2019) · Interspeech 2019 · doi:10.21437/interspeech.2019-2249 · Cited by 687
SUPERB: 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
XLS-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
Emotion Recognition from Speech Using Wav2vec 2.0 Embeddings
Pepino, L., Riera, P., & Ferrer, L. (2021) · Interspeech · doi:10.48550/arxiv.2104.03502 · Cited by 530
Attentive Statistics Pooling for Deep Speaker Embedding
Okabe, K., Koshinaka, T., & Shinoda, K. (2018) · Interspeech 2018 · doi:10.21437/interspeech.2018-993 · Cited by 514 · cited 2×
“Additionally, we pro-pose the Multi-Fusion Attentive (MFA) classifier based on the attentive statistics pooling layer [10].”
Method
Cited by
Deepfake Media Forensics: Status and Future Challenges
Amerini, I., Barni, M., Battiato, S., Bestagini, P., Boato, G., Bruni, V., Caldelli, R., De Natale, F., De Nicola, R., Guarnera, L., et al. (2025) · Journal of Imaging · doi:10.3390/jimaging11030073 · Cited by 113
Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead
Zhang, B., Cui, H., Nguyen, V., & Whitty, M. (2025) · Sensors · doi:10.3390/s25071989 · Cited by 72
Audio Deepfake Detection with Self-Supervised XLS-R and SLS Classifier
Zhang, Q., Wen, S., & Hu, T. (2024) · ACM International Conference on Multimedia (ACM MM) · doi:10.1145/3664647.3681345 · Cited by 86
XLSR-Mamba: A Dual-Column Bidirectional State Space Model for Spoofing Attack Detection
Xiao, Y., & Das, R. K. (2025) · IEEE Signal Processing Letters · doi:10.1109/lsp.2025.3547861 · Cited by 56
Anomaly Detection of Deepfake Audio Based on Real Audio Using Generative Adversarial Network Model
Song, D., Lee, N., Kim, J., & Choi, E. (2024) · IEEE Access · doi:10.1109/access.2024.3506973 · Cited by 24
Metadata
- Authors
- Guo, Y., Huang, H., Chen, X., Zhao, H., & Wang, Y.
- Affiliations
- Zhejiang University (CN)
- Venue
- ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
- Publisher
- IEEE
- Published
- 2024-04-14
- Pages
- 12702-12706
- ISSN
- 1520-6149 (ISSN-L)
- OpenAlex
- W4392902854
- Cited by
- 95
- References
- 27
- Citations by year
- 2026: 34 · 2025: 51 · 2024: 10
- FWCI
- 27.90
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- IEEE International Conference on Acoustics Speech and Signal Processing (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
- Physical Sciences › Computer Science › Signal Processing › Speech and Audio Processing
- Physical Sciences › Computer Science › Signal Processing › Music and Audio Processing
- Keywords
- audio deepfake detection, anti-spoofing, speaker verification, WavLM, self-supervised learning, voice conversion, speech synthesis
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 2026-09-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (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: Audio Deepfake Detection With Self-Supervised Wavlm And Multi-Fusion Attentive Classifier. https://app.cochat.ai/papers/10.1109/icassp48485.2024.10447923
- The paper
- Guo, Y., Huang, H., Chen, X., Zhao, H., & Wang, Y. (2024). Audio Deepfake Detection With Self-Supervised Wavlm And Multi-Fusion Attentive Classifier. ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). https://doi.org/10.1109/icassp48485.2024.10447923
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.