Elamur Research⌕
☰
home / papers / 2026 / When depth is redundant: Efficient transformer-based speech anti-spoofing

When depth is redundant: Efficient transformer-based speech anti-spoofing

Paper · 2026 · Hoan My Tran, Damien Lolive, Aghilas Sini et al.

Summary

The paper tackles speech deepfake detection. Here, transformer-based countermeasures score well on in-domain benchmarks but often fail on unseen, out-of-domain spoofing attacks. The authors studied how representations change across the layers of the transformer classifier placed on top of a pretrained XLS-R front end. They trained a linear probe on the final block and applied it unchanged to earlier layers. This showed that discriminative cues appear early and that adjacent layers, especially in shallow stacks, carry highly similar information. Extra depth therefore adds mostly redundancy and overfitting rather than robustness.

The method has two parts. An angular-distance regularizer pulls shallow and intermediate layer representations toward the final layer's task-specific representation, so early layers are actually used. Multi-head temporal latent attention (MTLA) replaces standard attention to cut memory and inference cost. Across several benchmarks, including multilingual and real-world data, the compact model matches or beats state-of-the-art systems with far fewer parameters. This matters because efficient, generalizable detectors are needed against fraud and impersonation.

Key points

  • Front end: XLS-R, a wav2vec 2.0-based multilingual SFM. It was pretrained on about 436,000 hours of unlabeled audio in 128 languages and has 24 transformer layers with hidden size 1024 and frames of roughly 20 ms.
  • Layer similarity is measured with the normalized angular distance of Gromov et al. (2025), applied to temporally pooled hidden states.
  • MTLA (Deng and Woodland, 2025) works in two steps:
    • it compresses features into a low-rank latent space, following the MLA approach;
    • it then merges adjacent time steps with hypernetwork-generated gates using stride s, which shrinks the KV cache by a factor of s.
  • Central finding: making the classifier deeper fails to consistently boost generalization to out-of-domain data and can even worsen overfitting. Aligning layers to the final representation improves robustness to unseen attacks.
  • Stated limitations:
    • the analysis covers only the classifier module, not the SFM;
    • SSM-type architectures are left for future work;
    • the alignment weight is a fixed hyperparameter that may need retuning for new domains.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🔬Centre National de la Recherche ScientifiqueParis · France168 works · 62,369 citations🎓Le Mans UniversitéLe Mans · France15 works · 31,097 citations🔬Institut de Recherche en Informatique et Systèmes AléatoiresRennes · France7 works · 1,944 citations🎓Université de RennesRennes · France3 works

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSignal Processing · 452 works

🧭 Topics 1Research subjects and languages the work is about.Efficient models & compression · 242 works

Cite

Cite as Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-Francois Marteau and David Guennec. 2026. When depth is redundant: Efficient transformer-based speech anti-spoofing. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.318

@inproceedings{tran2026when,
  title = {{When depth is redundant: Efficient transformer-based speech anti-spoofing}},
  author = {Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-Francois Marteau and David Guennec},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.318},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.318}
}

References

64 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. df arena: A leaderboard for speech deepfake detection models. arXiv preprint arXiv:2509.02859. — [arXiv:2509.02859](https://arxiv.org/abs/2509.02859) find
  2. Jiawei Du, I-Ming Lin, I-Hsiang Chiu, Xuanjun Chen, Haibin Wu, Wenze Ren, Yu Tsao, Hung-Yi Lee, and Jyh-Shing Roger Jang. 2024. Dfadd: The diffusion and flow-matching based audio deepfake dataset. In 2024 IEEE Spoken Language Technology Workshop (SLT), pages 921–928. find
  3. Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, Xu Tan, Yanqing Liu, Sheng Zhao, and Naoyuki Kanda. 2024. E2 tts: Embarrassingly easy fully non-autoregressive zero-shot tts. In 2024 IEEE Spoken Language Technology Workshop (… find
  4. Femi Folorunsho and Benedicta Frema Boamah. 2025. Deepfake technology and its impact: ethical considerations, societal disruptions, and security threats in ai-generated media. International journal of information technology and management information systems, 16(1):1060–1080. find
  5. Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Dan Roberts. 2025. The unreasonable ineffectiveness of the deeper layers. In The Thirteenth International Conference on Learning Representations.
  6. Albert Gu and Tri Dao. 2024. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling. find
  7. Yu Guan, Wu Guo, Jie Zhang, and Zhijun Zhang. 2025. Fusing multi-layer features of the pre-trained model with grouped cross attention for spoofing speech detection. In 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 601–606. find
  8. Yunqi Hao, Minqiang Xu, Yihao Chen, Yanyan Liu, Liang He, Lei Fang, and Lin Liu. 2025. Integrating spectro-temporal cross aggregation and multi-scale dynamic learning for audio deepfake detection. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)… find
  9. Wen Huang, Yanmei Gu, Zhiming Wang, Huijia Zhu, and Yanmin Qian. 2025. SpeechFake: A large-scale multilingual speech deepfake dataset incorporating cutting-edge generation methods. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pag… find
  10. Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, Detai Xin, Dongchao Yang, Eric Liu, Yichong Leng, Kaitao Song, Siliang Tang, Zhizheng Wu, Tao Qin, Xiangyang Li, Wei Ye, Shikun Zhang, Jiang Bian, Lei He, Jinyu Li, and sheng zhao. 2024. Naturalspeech 3: Zero-shot speech synthesis with factorized codec an… find
  11. Jee-weon Jung, Yihan Wu, Xin Wang, Ji-Hoon Kim, Soumi Maiti, Yuta Matsunaga, Hye-jin Shim, Jinchuan Tian, Nicholas Evans, Joon Son Chung, Wangyou Zhang, Seyun Um, Shinnosuke Takamichi, and Shinji Watanabe. 2025. Spoofceleb: Speech deepfake detection and sasv in the wild. IEEE Open Journal of Signal… find
  12. Taewoo Kim, Guisik Kim, Choongsang Cho, and Young Han Lee. 2025. Naturalness-Aware Curriculum Learning with Dynamic Temperature for Speech Deepfake Detection. In Interspeech 2025, pages 5318–5322. find
  13. Howard Lei and Eduardo Lopez. 2009. Mel, linear, and antimel frequency cepstral coefficients in broad phonetic regions for telephone speaker recognition. In Interspeech 2009, pages 2323–2326. find
  14. Xinfeng Li, Kai Li, Yifan Zheng, Chen Yan, Xiaoyu Ji, and Wenyuan Xu. 2024. Safeear: Content privacy-preserving audio deepfake detection. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, CCS ’24, page 3585–3599, New York, NY, USA. Association for Computing… find
  15. Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, and 1 others. 2024. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434. find
  16. Tianchi Liu, Duc-Tuan Truong, Rohan Kumar Das, Kong Aik Lee, and Haizhou Li. 2025. Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing. IEEE Transactions on Information Forensics and Security, 20:12005– 12018. find
  17. Xuechen Liu, Xin Wang, Md Sahidullah, Jose Patino, Héctor Delgado, Tomi Kinnunen, Massimiliano Todisco, Junichi Yamagishi, Nicholas Evans, Andreas Nautsch, and 1 others. 2023. find
  18. Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2507–2522. find
  19. Keith Ito and Linda Johnson. 2017. The lj speech dataset. https://keithito.com/ LJ-Speech-Dataset/. ↗
  20. Joao Medeiros. 2015. How intel gave stephen hawking a voice. Wired Retrieved from https://www.wired. com/2015/01/intel-gave-stephen-hawking-voice. ↗
  21. Jiachen Jiang, Jinxin Zhou, and Zhihui Zhu. 2025. Tracing representation progression: Analyzing and enhancing layer-wise similarity. In The Thirteenth International Conference on Learning Representations. find
  22. Nicolas Müller, Pavel Czempin, Franziska Diekmann, Adam Froghyar, and Konstantin Böttinger. 2022. Does audio deepfake detection generalize? In Interspeech 2022, pages 2783–2787. find
  23. Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger. 2024. Mlaad: The multi-language audio anti-spoofing dataset. In 2024 International Joint Conference on Neural Networks (IJCNN), pages 1–7. find
  24. Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: An asr corpus based on public domain audio books. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5206–5210. find
  25. Vardan Papyan, XY Han, and David L Donoho. 2020. Prevalence of neural collapse during the terminal phase of deep learning training. Proceedings of the National Academy of Sciences, 117(40):24652– 24663. find
  26. Ankita Pasad, Ju-Chieh Chou, and Karen Livescu. 2021. Layer-wise analysis of a self-supervised speech representation model. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), pages 914–921. find
  27. Ankita Pasad, Bowen Shi, and Karen Livescu. 2023. Comparative layer-wise analysis of self-supervised speech models. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. find
  28. Tuan Dat Phuong, Long-Vu Hoang, and Huy Dat Tran. 2025. Pushing the Performance of Synthetic Speech Detection with Kolmogorov-Arnold Networks and Self-Supervised Learning Models. In Interspeech 2025, pages 5633–5637. find
  29. Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. In Interspeech 2020, pages 2757–2761.
  30. Ricardo Reimao and Vassilios Tzerpos. 2019. For: A dataset for synthetic speech detection. In 2019 International Conference on Speech Technology and Human-Computer Dialogue (SpeD), pages 1–10. find
  31. Eros Rosello, Alejandro Gomez-Alanis, Angel M. Gomez, and Antonio Peinado. 2023. A conformer-based classifier for variable-length utterance processing in anti-spoofing. In Interspeech 2023, pages 5281–5285. find
  32. Peter Súkeník, Christoph H. Lampert, and Marco Mondelli. 2025. Neural collapse is globally optimal in deep regularized resnets and transformers. In The Thirty-ninth Annual Conference on Neural Information Processing Systems. find
  33. Chengzhe Sun, Shan Jia, Shuwei Hou, and Siwei Lyu. 2023. Ai-synthesized voice detection using neural vocoder artifacts. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 904–912. find
  34. Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, and Nicholas Evans. 2022a. Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP… find
  35. Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee weon Jung, Junichi Yamagishi, and Nicholas Evans. 2022b. Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation. In The Speaker and Language Recognition Workshop (Odyssey 2022), pages 112–119. find
  36. Pablo Andrés Tamayo Flórez, Rubén Manrique, and Bernardo Pereira Nunes. 2023. Habla: A dataset of latin american spanish accents for voice anti-spoofing. In Interspeech 2023, pages 1963–1967. find
  37. Xu Tan, Tao Qin, Frank Soong, and Tie-Yan Liu. 2021. A survey on neural speech synthesis. arXiv preprint arXiv:2106.15561. find
  38. Massimiliano Todisco, Héctor Delgado, and Nicholas Evans. 2017. Constant q cepstral coefficients: A spoofing countermeasure for automatic speaker verification. Computer Speech & Language, 45:516–535. find
  39. Massimiliano Todisco, Xin Wang, Ville Vestman, Md. Sahidullah, Héctor Delgado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi H. Kinnunen, and Kong Aik Lee. 2019. find
  40. Asvspoof 2019: Future horizons in spoofed and fake audio detection. In Interspeech 2019, pages 1008–1012. find
  41. Hoan My Tran, Damien Lolive, David Guennec, Aghilas Sini, Arnaud Delhay, and Pierre-François Marteau. 2025a. Kan you hear the truth? audio deepfake detection with kolmogorov–arnold networks. In 2025 33rd European Signal Processing Conference (EUSIPCO), pages 561–565. find
  42. Hoan My Tran, Damien Lolive, David Guennec, Aghilas Sini, Arnaud Delhay, and Pierre-François Marteau. 2025b. Leveraging SSL Speech Features and Mamba for Enhanced DeepFake Detection. In Interspeech 2025, pages 5323–5327. find
  43. Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-François Marteau, and David Guennec. 2025c. Multi-level ssl feature gating for audio deepfake detection. In Proceedings of the 33rd ACM International Conference on Multimedia, MM ’25, page 11766–11775, New York, NY, USA. Association fo… find
  44. Duc-Tuan Truong, Tianchi Liu, Junjie Li, Ruijie Tao, Kong Aik Lee, and Eng Siong Chng. 2025a. Addressing gradient misalignment in data-augmented training for robust speech deepfake detection. arXiv preprint arXiv:2509.20682. — [arXiv:2509.20682](https://arxiv.org/abs/2509.20682) find
  45. Duc-Tuan Truong, Tianchi Liu, Ruijie Tao, Junjie Li, Kong Aik Lee, and Eng Siong Chng. 2025b. find
  46. Qamo: Quality-aware multi-centroid one-class learning for speech deepfake detection. arXiv preprint arXiv:2509.20679. — [arXiv:2509.20679](https://arxiv.org/abs/2509.20679) find
  47. Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng. 2024. Temporal-channel modeling in multi-head self-attention for synthetic speech detection. In Interspeech 2024, pages 537–541. find
  48. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.
  49. Xin Wang, Héctor Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi Kinnunen, Nicholas Evans, Kong Aik Lee, Junichi Yamagishi, Myeonghun Jeong, Ge Zhu, Yongyi Zang, You Zhang, Soumi Maiti, Florian Lux, and 10 others. 2026. Asvspoof… find
  50. Xin Wang, Héctor Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi H. Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi. 2024. Asvspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale. In The A… find
  51. Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, and 1 others. 2020. find
  52. Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Computer Speech & Language, 64:101114. find
  53. Yang Xiao and Rohan Kumar Das. 2025. Xlsr-mamba: A dual-column bidirectional state space model for spoofing attack detection. IEEE Signal Processing Letters, 32:1276–1280. find
  54. Xi Xuan, Zimo Zhu, Wenxin Zhang, Yi-Cheng Lin, and Tomi Kinnunen. 2025. Fake-mamba: Real-time speech deepfake detection using bidirectional mamba as self-attention’s alternative. arXiv preprint arXiv:2508.09294. — [arXiv:2508.09294](https://arxiv.org/abs/2508.09294) find
  55. Junichi Yamagishi, Christophe Veaux, and Kirsten Mac-Donald. 2019. Cstr vctk corpus: English multi-speaker corpus for cstr voice cloning toolkit (version 0.92). The Rainbow Passage which the speakers read out can be found in the International Dialects of English Archive:(http://web. ku. edu/˜ idea/r… find
  56. Junichi Yamagishi, Xin Wang, Massimiliano Todisco, Md Sahidullah, Jose Patino, Andreas Nautsch, Xuechen Liu, Kong Aik Lee, Tomi Kinnunen, Nicholas Evans, and Héctor Delgado. 2021. find
  57. Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection. In 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, pages 47–54. find
  58. Mingru Yang, Yanmei Gu, Qianhua He, Yanxiong Li, Peirong Zhang, Yongqiang Chen, Zhiming Wang, Huijia Zhu, Jian Liu, and Weiqiang Wang. 2025a. Generalizable Audio Deepfake Detection via Hierarchical Structure Learning and Feature Whitening in Poincaré sphere. In Interspeech 2025, pages 2255– 2259. find
  59. Mingru Yang, Yanmei Gu, Qianhua He, Peirong Zhang, Haolin He, Zhiming Wang, Huijia Zhu, Jian Liu, and Weiqiang Wang. 2025b. Generalizable audio deepfake detection via risk-aware style alignment and structural empirical risk minimization. In Proceedings of the 33rd ACM International Conference on Mul… find
  60. Jiangyan Yi, Jianhua Tao, Ruibo Fu, Xinrui Yan, Chenglong Wang, Tao Wang, Chu Yuan Zhang, Xiaohui Zhang, Yan Zhao, Yong Ren, and 1 others. 2023. find
  61. Add 2023: the second audio deepfake detection challenge. arXiv preprint arXiv:2305.13774. — [arXiv:2305.13774](https://arxiv.org/abs/2305.13774) find
  62. Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J. Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. 2019. Libritts: A corpus derived from librispeech for text-to-speech. In Interspeech 2019, pages 1526–1530.
  63. Qishan Zhang, Shuangbing Wen, and Tao Hu. 2024. Audio deepfake detection with self-supervised XLS-r and SLS classifier. In ACM Multimedia 2024. find
  64. Zhenyu Zhang, Yewei Gu, Xiaowei Yi, and Xianfeng Zhao. 2021. Fmfcc-a: a challenging mandarin dataset for synthetic speech detection. In International Workshop on Digital Watermarking, pages 117–131. Springer. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tran-2026-when-depth-is-redundant-efficient/