The DOI matches its Crossref record; 34 of the 34 references it cites were found in a registry or an index; 3 of the 9 sources its claims cite were checked for support.
The Vicomtech Audio Deepfake Detection System Based on Wav2vec2 for the 2022 ADD Challenge
Open accessAbstract
This paper describes our submitted systems to the 2022 ADD challenge withing the tracks 1 and 2. Our approach is based on the combination of a pre-trained wav2vec2 feature extractor and a downstream classifier to detect spoofed audio. This method exploits the contextualized speech representations at the different transformer layers to fully capture discriminative information. Furthermore, the classification model is adapted to the application scenario using different data augmentation techniques. We evaluate our system for audio synthesis detection in both the ASVspoof 2021 and the 2022 ADD challenges, showing its robustness and good performance in realistic challenging environments such as telephonic and audio codec systems, noisy audio, and partial deepfakes.
Overview
CoChat Summary
TL;DR
The paper presents Vicomtech’s wav2vec2-based audio deepfake detection systems for the 2022 ADD Challenge tracks 1 and 2, using frozen W2V2 representations from multiple transformer layers with a lightweight downstream classifier. The systems ranked first in track 1 and fourth in track 2, and also showed competitive EERs on ASVspoof 2021 LA and DF.
Key findings
- In ADD 2022, the best submitted system achieved 21.71% EER on track 1 and 16.59% EER on track 2, ranking first and fourth among participants, respectively.
- Using adaptation data with the train set reduced EER substantially, e.g. XLS-128 on track 1 from 32.20% to 22.62%, and on track 2 from 45.88% to 30.35%.
- Low-pass FIR augmentation reduced EER by about 1% in both ADD tracks; for track 2, partial-fake augmentation further improved the best result to 16.59% EER.
- On ASVspoof 2021, the proposed system reached 3.54% EER on LA and 4.98% EER on DF, outperforming other methods on DF and achieving competitive LA results despite being a single system rather than an ensemble.
Methods
- The system uses pre-trained wav2vec2 Large models (300M parameters, XLS-53 and XLS-128) as frozen feature extractors, feeding hidden representations from different transformer layers into a downstream classifier with trainable layer weights, temporal normalization, feedforward layers, attentive statistical pooling, and a cosine scoring layer trained with one-class softmax loss.
- ADD 2022 experiments used the challenge train/dev sets of about 28K utterances each, adaptation sets of about 1K utterances, and test sets of about 100K unlabeled utterances; ASVspoof 2021 experiments used ASVspoof 2019 LA train/dev sets of about 25K utterances and evaluated on the 2021 LA and DF partitions.
- Data augmentation included narrowband and wideband low-pass FIR filters applied to raw waveforms, plus for ADD track 2 a partial-fake generation strategy that randomly selected 20% of genuine utterances per epoch and overlapped a variable-duration segment from another utterance. Training used Adam with default learning rate, dropout 0.2, mini-batch 8, parameter updates every 8 iterations, and early stopping after 10 epochs without dev improvement.
Limitations
- The paper does not list explicit limitations; it states as future work testing other self-supervised models and additional data augmentation techniques.
- Evident from the text: all systems were trained under closed conditions using only the corresponding challenge train data, with ADD adaptation sets being small (~1K utterances), so broader cross-dataset generalization is not established.
- Evident from the text: the W2V2 parameters were frozen during training, and the track 2 partial-fake augmentation is a heuristic generation procedure rather than a learned or exhaustively validated augmentation strategy.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 1 Not valid · 6 Not checked · 26 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“Nevertheless, previous works have shown that for other tasks, for example, speaker verification or emotion recognition, more discriminative information can be obtained from the first or intermediate layers [12, 13].”
What the sources say
- Supportedcites [12] [12] Z. Chen, S. Chen, Y. Wu, Y. Qian, C. Wang, S. Liu, Y. Qian, and M. Zeng, 'Large-scale self-supervised speech representation learning for automatic speaker verification,' arXiv preprint arXiv:2110.05777 , 2021. - source
Section 5.3 (p. 4) finds that lower-layer representations are more discriminative than higher-layer representations for speaker verification. Section 3.1 (p. 2) identifies these layers as part of the pretrained models’ Transformer encoders. This source supports the speaker-verification example; the separate emotion-recognition example is not attributed specifically to it. (LLM, conf=0.93)
- Supportedcites [13] [13] L. Pepino, P. Riera, and L. Ferrer, 'Emotion Recognition from Speech Using wav2vec 2.0 Embeddings,' in Proc. Interspeech 2021 , 2021, pp. 3400-3404. - source
The Results and discussion section reports that middle wav2vec 2.0 layers receive larger learned weights and that including intermediate layers improves emotion-recognition performance over using encoder outputs alone. This supports the emotion-recognition part attributed to this paper; the separate speaker-verification example need not be established by this citation. (LLM, conf=0.88)
plainstates 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
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
“We evaluate our system for audio synthesis detection in both the ASVspoof 2021 and the 2022 ADD challenges, showing its robustness and good performance in realistic challenging environments such as telephonic and audio codec systems, noisy audio, and partial deepfakes.”
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.
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
“Furthermore, the FIR-based augmentations help to improve the discrimination capabilities of our proposal.”
plainstates a cause- 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
“The XLS-128 model shows better results in general, but the main improvements come from the use of adaptation data along with the train set.”
plainstates an association- 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
“This strategy reduced the EER by about 1% in both tracks and combined with the previous ones, it achieved the best results.”
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.
10 more claims
Create a free account to see all 35 claims and every source verdict.
- Probably not a substantial claim. 10
- It may state two separate claims as one. 13
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
“The study of Martín-Doñas et al. [126] explored both XLSR and XLS-R as feature extractors for audio deepfake detection. By integrating hidden representations from different Transformer layers, they achieved remarkable performance on the ASVspoof 2021 and ADD 2022 datasets”
From Audio Deepfake Detection: What Has Been Achieved and What Lies Ahead (2025) · cites this paper as [126]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Previous work [5, 12, 13] has demonstrated that the intermediate feature of self-supervised model contains certain discriminative information.”
From Audio Deepfake Detection With Self-Supervised WavLM And Multi-Fusion Attentive Classifier (2024) · cites this paper as [5]
Citations
Cites
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946
Instance Normalization: The Missing Ingredient for Fast Stylization
Ulyanov, D., Vedaldi, A., & Lempitsky, V. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1607.08022 · Cited by 2,775
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
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
Spoofing and countermeasures for speaker verification: A survey
Wu, Z., Evans, N., Kinnunen, T., Yamagishi, J., Alegre, F., & Li, H. (2015) · Speech Communication · doi:10.1016/j.specom.2014.10.005 · Cited by 601
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
ASVspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge
Wu, Z., Kinnunen, T., Evans, N., Yamagishi, J., Hanilçi, C., Sahidullah, M., & Sizov, A. (2015) · Interspeech 2015 · doi:10.21437/interspeech.2015-462 · Cited by 484
Cited by
ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild
Liu, X., Wang, X., Sahidullah, M., Patino, J., Delgado, H., Kinnunen, T., Todisco, M., Yamagishi, J., Evans, N., Nautsch, A., et al. (2023) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2023.3285283 · Cited by 312
A Survey on Speech Deepfake Detection
Li, M., Ahmadiadli, Y., & Zhang, X.-P. (2025) · ACM Computing Surveys · doi:10.1145/3714458 · Cited by 101
Zhang, L., Wang, X., Cooper, E., Evans, N., & Yamagishi, J. (2023) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2022.3233236 · Cited by 100
Audio Deepfake Detection With Self-Supervised WavLM And Multi-Fusion Attentive Classifier
Guo, Y., Huang, H., Chen, X., Zhao, H., & Wang, Y. (2024) · ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp48485.2024.10447923 · Cited by 95
MLAAD: The Multi-Language Audio Anti-Spoofing Dataset
Müller, N. M., Kawa, P., Choong, W. H., Casanova, E., Gölge, E., Müller, T., Syga, P., Sperl, P., & Böttinger, K. (2024) · 2024 International Joint Conference on Neural Networks (IJCNN) · doi:10.1109/ijcnn60899.2024.10650962 · Cited by 73
Metadata
- Authors
- Martin-Donas, J. M., & Alvarez, A.
- Affiliations
- Vicomtech Foundation, Basque Research and Technology Alliance (BRTA),Donostia – San Sebastián,Spain,20009
- Venue
- ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
- Publisher
- IEEE
- Published
- 2022-05-23
- Pages
- 9241-9245
- ISSN
- 1520-6149 (ISSN-L)
- OpenAlex
- W4226264925
- Cited by
- 100
- References
- 31
- Citations by year
- 2026: 16 · 2025: 25 · 2024: 36 · 2023: 18 · 2022: 5
- FWCI
- 11.60
- 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, speech spoofing detection, wav2vec 2.0, transformer layers, data augmentation, ASVspoof 2021, noisy audio
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.
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 (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: The Vicomtech Audio Deepfake Detection System Based on Wav2vec2 for the 2022 ADD Challenge. https://app.cochat.ai/papers/10.1109/icassp43922.2022.9747768
- The paper
- Martin-Donas, J. M., & Alvarez, A. (2022). The Vicomtech Audio Deepfake Detection System Based on Wav2vec2 for the 2022 ADD Challenge. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). https://doi.org/10.1109/icassp43922.2022.9747768
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.