The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index; 14 of the 19 sources its claims cite were checked for support.
Robust DeepFake Audio Detection via an Improved NeXt-TDNN with Multi-Fused Self-Supervised Learning Features
Open accessAbstract
Deepfake audio refers to speech that has been synthetically generated or altered through advanced neural network techniques, often with a degree of realism sufficient to convincingly imitate genuine human voices. As these manipulations become increasingly indistinguishable from authentic recordings, they present significant threats to security, undermine media integrity, and challenge the reliability of digital authentication systems. In this study, a robust detection framework is proposed, which leverages the power of self-supervised learning (SSL) and attention-based modeling to identify deepfake audio samples. Specifically, audio features are extracted from input speech using two powerful pretrained SSL models: HuBERT-Large and WavLM-Large. These distinctive features are then integrated through an Attentional Multi-Feature Fusion (AMFF) mechanism. The fused features are subsequently classified using a NeXt-Time Delay Neural Network (NeXt-TDNN) model enhanced with Efficient Channel Attention (ECA), enabling improved temporal and channel-wise feature discrimination. Experimental results show that the proposed method achieves a 0.42% EER and 0.01 min-tDCF on ASVspoof 2019 LA, a 1.01% EER on ASVspoof 2019 PA, and a pooled 6.56% EER on the cross-channel ASVspoof 2021 LA evaluation, thus highlighting its effectiveness for real-world deepfake detection scenarios. Furthermore, on the ASVspoof 5 dataset, the method achieved a 7.23% EER, outperforming strong baselines and demonstrating strong generalization ability. Moreover, the macro-averaged F1-score of 96.01% and balanced accuracy of 99.06% were obtained on the ASVspoof 2019 LA dataset, while the proposed method achieved a macro-averaged F1-score of 98.70% and balanced accuracy of 98.90% on the ASVspoof 2019 PA dataset. On the highly challenging ASVspoof 5 dataset, which includes crowdsourced, non-studio-quality audio, and novel adversarial attacks, the proposed method achieves macro-averaged metrics exceeding 92%, with a precision of 92.07%, a recall of 92.63%, an F1-measure of 92.35%, and a balanced accuracy of 92.63%.
Overview
CoChat Summary
TL;DR
The paper proposes a deepfake audio detector that fuses 8th-layer HuBERT-Large and WavLM-Large embeddings via Attentional Multi-Feature Fusion and classifies them with an ECA-enhanced lightweight NeXt-TDNN. It reports state-of-the-art EER/min-tDCF on ASVspoof 2019 LA and competitive cross-dataset results on ASVspoof 2019 PA, 2021 LA, and ASVspoof 5.
Key findings
- On ASVspoof 2019 LA, the proposed method achieves 0.42% EER and 0.01 min-tDCF, reported as outperforming the previous best AASIST family by roughly 49–50% relative; Table 12 lists the best prior system at 0.80% EER and 0.02 min-tDCF, with relative margins of 47% and 50%.
- On ASVspoof 2019 PA, it achieves 1.01% EER and 0.03 min-tDCF pooled, with per-condition EER from 0.48% to 1.36% and min-tDCF from 0.0134 to 0.0397 across nine device-distance conditions.
- On ASVspoof 2021 LA, it reaches a pooled 6.56% EER, reported as roughly 31% relative reduction over RawNet2 (9.50%) and almost 58% over CQCC-GMM (15.62%); on ASVspoof 5, it achieves 7.23% EER with precision 92.07%, recall 92.63%, F1 92.35%, and balanced accuracy 92.63%.
- Ablation: AMFF fusion yields 0.42% EER and 0.012 min-tDCF, versus 0.89% for feature concatenation and 1.72% for mean score fusion; the ECA-enhanced NeXt-TDNN-l improves over NeXt-TDNN-l, e.g., WavLM 0.64% vs 0.74% EER and HuBERT 0.63% vs 0.86% EER.
Methods
- Design: raw waveform is processed by pretrained HuBERT-Large and WavLM-Large, with 1024-dimensional embeddings extracted from the 8th Transformer layer; these are concatenated and fused by the AMFF Parallel-Supplementary Weighted Fusion strategy into a 1024-dimensional vector, then classified by an improved NeXt-TDNN-l with Efficient Channel Attention and AM-Softmax (scale 40, margin 0.3).
- Data: ASVspoof 2019 LA and PA, ASVspoof 2021 LA (evaluation-only), and ASVspoof 5 Track 1. Training used PyTorch 1.8, Adam optimizer with learning rate 1 × 10⁻⁴, batch size 32, 100 epochs, and the checkpoint with lowest validation EER; experiments ran on an Intel i7 CPU, 64 GB RAM, and GeForce RTX 3060.
- Metrics: EER and minimum tandem Detection Cost Function (min-tDCF), plus macro-averaged precision, recall, F1, and balanced accuracy because of class imbalance.
Limitations
- Stated: future work is needed to extend robustness to emerging neural vocoders, streaming codecs, and adversarial manipulations, and to integrate the detector more tightly with speaker verification back-ends.
- Evident: performance drops on ASVspoof 5 (7.23% EER) and ASVspoof 2021 LA (6.56% pooled EER) relative to ASVspoof 2019 LA (0.42% EER); the paper also notes degradation on vocoder-based attack A10 and lower precision/accuracy (0.5995) on attacks A13 and A17 despite perfect recall.
- Evident: ASVspoof 2021 LA is evaluation-only, posing a cross-corpus generalization scenario, and ASVspoof 5 includes crowdsourced, non-studio-quality audio and novel adversarial attacks.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
12 Valid · 4 Not valid · 4 Not checked · 73 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“this design poses a more challenging cross-corpus generalization scenario, as systems trained on earlier corpora must transfer to unseen spoofing attacks encountered only at test time”
What the sources say
- Partly supportedcites [26] Yamagishi, J., Wang, X., Todisco, M., Sahidullah, M., Patino, J., Nautsch, A., Liu, X., Lee, K.A., Kinnunen, T., and Evans, N. (2021, January 16). ASVspoof 2021: Accelerating progress in spoofed and deepfake speech detection. Proceedings of the ASVspoof 2021 Workshop-Automatic Speaker Verification and Spoofing Coutermeasures Challenge, Online. sourceExperimental — accuracy not yet measured
The Database section (p. 1) says participants were expected to use ASVspoof 2019 training and development data for new 2021 evaluation partitions, and calls the 2021 LA data intentionally more challenging. The Introduction (p. 1) attributes added LA difficulty to telephony encoding and transmission. Our copy does not say that the LA task introduced spoofing attacks encountered only at test time, so that specific explanation is unsupported. (LLM, conf=0.85)
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.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Kheir et al. (2025) conducted a comprehensive layer-wise analysis of HuBERT and WavLM, and reported that lower-to-mid layers consistently provide the most discriminative features across diverse datasets and attack scenarios [3].”
What the sources say
- Partly supportedcites [3] Kheir, Y.E., Samih, Y., Maharjan, S., Polzehl, T., and Möller, S. (2025). Comprehensive layer-wise analysis of ssl models for audio deepfake detection. arXiv. sourceExperimental — accuracy not yet measured
Section 5.1 (p. 7) reports that lower layers make the highest average contributions in large SSL models, including HuBERT-Large and WavLM-Large. But it says lower layers contribute significantly across *most* datasets, not that lower-to-mid layers are consistently the most discriminative in every scenario. Section 5.2 (p. 8) identifies PartialSpoof as an exception, where upper-middle layers contribute most. (LLM, conf=0.86)
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.
- Partly 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 demonstrate that the proposed system outperforms current state-of-the-art baselines on the ASVspoof 2019 LA, ASVspoof 2019 PA, and ASVspoof 2021 LA benchmarks.”
boosted- 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
“Extensive experiments on several public deepfake audio datasets demonstrate that our proposed method achieves state-of-the-art detection performance”
boosted- 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
“a pooled 6.56% EER on the cross-channel ASVspoof 2021 LA evaluation”
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.
36 more claims
Create a free account to see all 93 claims and every source verdict.
- Probably not a substantial claim. 11
- It may state two separate claims as one. 46
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.3. Ablation Studies
“Substituting the backbone with NeXt-TDNN (channel width C = 256, depthwise-conv kernel K = 7, block repetition B = 3) lowers the EER to 0.96% and nearly halves the t-DCF.”
§ 4.3. Ablation Studies
“| WavLM | NeXt-TDNN | 7.1M | 1.64 | 0.029 |”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.5. Performance Evaluation and Comparative Analysis on the ASVspoof 2019 PA Dataset
“| Proposed method | AMFF fused SSL features + Improved NeXt-TDNN | 0.42 | 0.019 |”
§ 4.3. Ablation Studies
“| Attentional Multi-Feature Fusion | HuBERT and WavLM | 0.42 | 0.012 |”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ 4.4. Performance Evaluation and Comparative Analysis on the ASVspoof 2019 LA Dataset
“it exhibits pronounced degradation on the more challenging vocoder-based attack A10 (4.39% EER, 0.052 min t-DCF).”
§ 4.4. Performance Evaluation and Comparative Analysis on the ASVspoof 2019 LA Dataset
“| A10 | 0.75 | 0.020 | 1.40 | 0.040 | 1.89 | 0.022 |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.3. Ablation Studies
“the light variant, NeXt-TDNN-l, is evaluated, which trims the model to 5.9M parameters yet still surpasses the baseline with an EER of 0.82%. The improved NeXt-TDNN-l with an ECA layer adds only three additional parameters but delivers the best overall performance-an EER of 0.64% while retaining the same minimum t-DCF of 0.020.”
§ 4.3. Ablation Studies
“| WavLM | NeXt-TDNN- l | 5.9M | 0.74 | 0.023 | | WavLM | NeXt-TDNN- l with ECA layer | 5.9 M+3 | 0.64 | 0.019 |”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.4. Performance Evaluation and Comparative Analysis on the ASVspoof 2019 LA Dataset
“In general, WavLM delivers the lowest EER on the majority of attacks (e.g., 0% for A09 and A13, and ≤ 0.08% for A07, A08, A12, A14-A16)”
§ 4.4. Performance Evaluation and Comparative Analysis on the ASVspoof 2019 LA Dataset
“| A10 | 0.75 | 0.020 | 1.40 | 0.040 | 1.89 | 0.022 | | A11 | 0.19 | 0.005 | 0.47 | 0.012 | 0.18 | 0.005 | | A12 | 0.12 | 0.003 | 0.06 | 0.001 | 0.04 | 0.001 |”
1 more finding
Create a free account to see every finding of this check.
Citations
Cites
ECA-Net: Efficient Channel Attention for Deep Convolutional Neural Networks
Wang, Q., Wu, B., Zhu, P., Li, P., Zuo, W., & Hu, Q. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.01155 · Cited by 9,641
A systematic analysis of performance measures for classification tasks
Sokolova, M., & Lapalme, G. (2009) · Information Processing & Management · doi:10.1016/j.ipm.2009.03.002 · Cited by 7,515
HuBERT: 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
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
The Balanced Accuracy and Its Posterior Distribution
Brodersen, K. H., Ong, C. S., Stephan, K. E., & Buhmann, J. M. (2010) · 2010 20th International Conference on Pattern Recognition · doi:10.1109/icpr.2010.764 · Cited by 1,816
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
Constant Q Cepstral Coefficients: A Spoofing Countermeasure for Automatic Speaker Verification
Todisco, M., Delgado, H., & Evans, N. (2017) · Computer Speech & Language · doi:10.1016/j.csl.2017.01.001 · Cited by 424
End-to-End anti-spoofing with RawNet2
Tak, H., Patino, J., Todisco, M., Nautsch, A., Evans, N., & Larcher, A. (2021) · ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp39728.2021.9414234 · Cited by 402
AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks
Jung, J.-W., Heo, H.-S., Tak, H., Shim, H.-J., Chung, J. S., Lee, B.-J., Yu, H.-J., & Evans, N. (2022) · ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp43922.2022.9747766 · Cited by 399
Research works printed but not linked: 1
- Wang, Z., Cui, S., Kang, X., Sun, W., and Li, Z. (2020, January 7–10). Densely connected convolutional network for audio spoofing detection. Proceedings of the 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Auckland, New Zealand.
Cited by
A multi-level and iterative feature engineering framework for deepfake audio detection
Çelik, B., & Akbal, A. (2025) · Expert Systems with Applications · doi:10.1016/j.eswa.2025.130722 · Cited by 5
Multi-Speaker Conversational Audio Deepfake: Taxonomy, Dataset and Pilot Study
Ahmed, A., Janeja, V. P., & Purushotham, S. (2025) · IEEE ... International Conference on Data Mining workshops · doi:10.1109/icdmw69685.2025.00163 · Cited by 3
“…• We plan to expand our benchmark comparisons to include recent transformer-based deepfake detection models such as those based on Audio Spectrogram Transformer (AST) [43]–[45] or Conformer architectures [46] and explore multi-modal approaches that combine acoustic and linguistic features.”
Deepfakes: A comprehensive survey on techniques, challenges and future directions
Usmani, S., Kumar, S., & Sadhya, D. (2026) · Computers & Electrical Engineering · doi:10.1016/j.compeleceng.2026.111210 · Cited by 3
Multi-Scale Self-Supervised Learning for Efficient Audio Deepfake Detection
Wani, T. M., & Amerini, I. (2025) · IEEE Signal Processing Letters · doi:10.1109/lsp.2025.3634032 · Cited by 2
“HuBert+WavLM [7] further illustrates this limitation, relying on computationally expensive ensemble architectures requiring multiple large models while lacking interpretability through the absence of handcrafted acoustic features.”
The Invisible War: Countering Adversarial Evasion and Anti-Forensic Masks in Synthetic Media
Jahangeer, G. S. B., Tupe, B. S., & Ariffin, N. H. M. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-2600-2278-8.ch009
Metadata
- Authors
- Tahaoglu, G.
- Affiliations
- Department of Computer Engineering, Karadeniz Technical University, Trabzon 61080, Türkiye
- ORCID
- Tahaoglu, Gul
- Venue
- Applied Sciences
- Publisher
- MDPI AG
- Published
- 2025-09-03
- Volume
- 15
- Issue
- 17
- Pages
- 9685
- ISSN
- 2076-3417 (ISSN-L)
- OpenAlex
- W4413991433
- Cited by
- 11
- References
- 43
- Citations by year
- 2026: 8 · 2025: 3
- FWCI
- 4.41
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Applied Sciences (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- 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 › Computer Vision and Pattern Recognition › Digital Media Forensic Detection
- Keywords
- audio deepfake detection, speech spoofing detection, self-supervised learning, attention-based feature fusion, HuBERT, WavLM, efficient channel attention, ASVspoof, adversarial attacks
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 2026-09-30, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-09-30 (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: Robust DeepFake Audio Detection via an Improved NeXt-TDNN with Multi-Fused Self-Supervised Learning Features. https://app.cochat.ai/papers/10.3390/app15179685
- The paper
- Tahaoglu, G. (2025). Robust DeepFake Audio Detection via an Improved NeXt-TDNN with Multi-Fused Self-Supervised Learning Features. Applied Sciences. https://doi.org/10.3390/app15179685
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.