Elamur Research⌕
☰
home / papers / 2026 / Bridging Attribution and Open-Set Detection using Graph-Augmented Instance Learning in Synthetic Speech

Bridging Attribution and Open-Set Detection using Graph-Augmented Instance Learning in Synthetic Speech

Summary

The authors combine two tasks in the forensic analysis of synthetic speech. The first is attribution, that is, identifying which TTS system generated a recording. The second is open-set recognition: detecting audio from a generator that was not seen during training. For this, they propose the SIGNAL framework. First, frozen speech foundation models (SFMs) produce utterance-level embeddings. Next, two branches run in parallel. An attention-based GNN (GAT) reasons over a graph of generator-class prototypes built for the specific query. A KNN assesses local instance similarity. An ensemble head combines the two branches, and a confidence threshold assigns the sample to a known generator or flags it as unknown. The system was trained and evaluated on DiffSSD, which contains real speech and output from open-source and commercial diffusion-based TTS systems. Transfer was tested in a zero-shot setting on the singing-voice SingFake. The hybrid consistently outperforms standalone KNN and GNN as well as the FCN/CNN baselines, and Audio-MAMBA embeddings give the best results. According to the authors, this is the first work to combine graph learning with open-set detection for tracing the source of synthetic speech.

Key points

  • Nine representations were compared: Whisper (base), WavLM, UniSpeech-SAT, wav2vec 2.0, x-vector, ECAPA-TDNN, and Audio-MAMBA in tiny, small and base variants (4.8M, 17.9M and 69.3M parameters). Embeddings were obtained by mean pooling of the last hidden state.
  • On SingFake with Mamba-B, the baseline classifiers achieve an F1 of 82.63% and an EER of 17.97% on the ID/OOD split. SIGNAL raises F1 to 90.99% and lowers the EER on the open ID/OOD split to 7.92%. With Mamba-S, the hybrid lowers the EER to 10.38%.
  • An ablation of the threshold τ over the range 0.1–0.9 on four splits (ID and ID/OOD for DiffSSD and SingFake) showed that τ = 0.5 gives the best EER balance.
  • The authors acknowledge limitations: the system only flags a sample as "unknown" and does not distinguish between multiple unknown generators. It also depends on a fixed threshold and does not explain its decisions.
  • The code and resources are publicly available: github.com/Helixometry/SIGNAL.

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

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-long.250

@inproceedings{akhtar2026bridging,
  title = {{Bridging Attribution and Open-Set Detection using Graph-Augmented Instance Learning in Synthetic Speech}},
  author = {Akhtar, Mohd Mujtaba and Girish and Sheth, Farhan and Singh, Muskaan},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-long.250},
  url = {https://doi.org/10.18653/v1/2026.eacl-long.250}
}

References

30 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449–12460. find our card
  2. Kratika Bhagtani, Amit Kumar Singh Yadav, Paolo Bestagini, and Edward J Delp. 2025. Diffssd: A diffusion-based dataset for speech forensics. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. find
  3. Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, and 1 others. 2022. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505–1518. find
  4. Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, and Xiangzhan Yu. 2021. Unispeech-sat: Universal speech representation learning with speaker aware pre-training. find
  5. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6152– 6156. find
  6. Erica Cooper, Cheng-I Lai, Yusuke Yasuda, Fuming Fang, Xin Wang, Nanxin Chen, and Junichi Yamagishi. 2020. Zero-shot multi-speaker text-to-speech with state-of-the-art neural speaker embeddings. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pa… find
  7. Brecht Desplanques, Jenthe Thienpondt, and Kris De-muynck. 2020. Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. In Interspeech 2020, pages 3830–3834. find
  8. Falih Gozi Febrinanto, Kristen Moore, Chandra Thapa, Jiangang Ma, Vidya Saikrishna, and Feng Xia. 2025. Vision graph non-contrastive learning for audio deepfake detection with limited labels. arXiv preprint arXiv:2501.04942. — [arXiv:2501.04942](https://arxiv.org/abs/2501.04942) find
  9. Albert Gu and Tri Dao. 2024. Mamba: Linear-time sequence modeling with selective state spaces. In First Conference on Language Modeling. find
  10. Yinlin Guo, Haofan Huang, Xi Chen, He Zhao, and Yuehai Wang. 2024. Audio deepfake detection with self-supervised wavlm and multi-fusion attentive classifier. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 12702–12706. IEEE. find
  11. Lian Huang and Chi-Man Pun. 2024. Self-attention and hybrid features for replay and deep-fake audio detection. arXiv preprint arXiv:2401.05614. — [arXiv:2401.05614](https://arxiv.org/abs/2401.05614) find
  12. Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, Detai Xin, Dongchao Yang, Yanqing Liu, Yichong Leng, Kaitao Song, Siliang Tang, Zhizheng Wu, Tao Qin, Xiang-Yang Li, Wei Ye, Shikun Zhang, Jiang Bian, Lei He, Jinyu Li, and Sheng Zhao. 2024. Naturalspeech 3: zero-shot speech synthesis with factorized code… find
  13. Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2021. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations. find
  14. Matthew Le, Apoorv Vyas, Bowen Shi, Brian Karrer, Leda Sari, Rashel Moritz, Mary Williamson, Vimal Manohar, Yossi Adi, Jay Mahadeokar, and 1 others. 2023. Voicebox: Text-guided multilingual universal speech generation at scale. Advances in neural information processing systems, 36:14005–14034. find
  15. Naihan Li, Shujie Liu, Yanqing Liu, Sheng Zhao, and Ming Liu. 2019. Neural speech synthesis with transformer network. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 6706–6713. find
  16. Jagabandhu Mishra, Manasi Chhibber, Hye-jin Shim, and Tomi H Kinnunen. 2025. Towards explainable spoofed speech attribution and detection: A probabilistic approach for characterizing speech synthesizer components. Computer Speech & Language, page 101840. find
  17. Orchid Chetia Phukan, Girish, Mohd Mujtaba Akhtar, Swarup Ranjan Behera, Priyabrata Mallick, Pailla Balakrishna Reddy, Arun Balaji Buduru, and Rajesh Sharma. 2025. Towards Source Attribution of Singing Voice Deepfake with Multimodal Foundation Models. In Interspeech 2025, pages 1673–1677. find
  18. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492–28518. PMLR. find our card
  19. Zhizheng Wu, Tomi Kinnunen, Nicholas Evans, Junichi Yamagishi, Cemal Hanilçi, Md Sahidullah, and Aleksandr Sizov. 2015. find
  20. Asvspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge. In INTERSPEECH 2015 16th Annual Conference of the International Speech Communication Association, pages 2037–2041. International Speech Communication Association. find
  21. Sarthak Yadav and Zheng-Hua Tan. 2024. Audio mamba: Selective state spaces for self-supervised audio representations. In Interspeech 2024, pages 552–556. find
  22. Jiangyan Yi, Chenglong Wang, Jianhua Tao, Xiaohui Zhang, Chu Yuan Zhang, and Yan Zhao. 2023. Audio deepfake detection: A survey. ArXiv, abs/2308.14970. — [arXiv:2308.14970](https://arxiv.org/abs/2308.14970) find
  23. Yongyi Zang, You Zhang, Mojtaba Heydari, and Zhiyao Duan. 2024. Singfake: Singing voice deepfake detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 12156–12160. IEEE. find
  24. Yi Ren, Chenxu Hu, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. 2021. Fastspeech 2: Fast and high-quality end-to-end text to speech. In International Conference on Learning Representations. find
  25. Hyun-seo Shin, Jungwoo Heo, Ju-ho Kim, Chan-yeong Lim, Wonbin Kim, and Ha-Jin Yu. 2024. Hm-conformer: A conformer-based audio deepfake detection system with hierarchical pooling and multi-level classification token aggregation methods. In ICASSP 2024-2024 IEEE International Conference on Acoustics,… find
  26. David Snyder, Daniel Garcia-Romero, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur. 2018. X-vectors: Robust dnn embeddings for speaker recognition. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5329–5333. find
  27. Adriana Stan, David Combei, Dan Oneata, and Horia Cucu. 2025. TADA: Training-free Attribution and Out-of-Domain Detection of Audio Deepfakes. In Interspeech 2025, pages 1543–1547. find
  28. Yun Tan, Xiaoqian Weng, and Jiangzhang Zhu. 2025. Dual-channel spoofed speech detection based on graph attention networks. Symmetry, 17(5):641. find
  29. 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
  30. Asvspoof 2019: Future horizons in spoofed and fake audio detection. In Interspeech 2019, pages 1008–1012. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/akhtar-2026-bridging-attribution-open-set-detection/