A benchmark for studying neural codec–based deepfake speech detection.
Published in IEEE Transactions on Audio, Speech and Language Processing, vol. 34, pp. 2929–2944, 2026.
🔗 Audio demo · Paper (IEEE) · arXiv · Dataset (Hugging Face)
CodecFake+ is a large-scale benchmark designed to evaluate the detection and generalization capabilities of models across diverse neural audio codec–generated deepfakes. It enables controlled, taxonomy-aware analysis for codec-based deepfake research.
| Subset | Role | Bona fide | Spoof | Models | Paper | Status |
|---|---|---|---|---|---|---|
| CoRS | Train / Val / Eval | 44,455 | 1,378,105 | 31 codecs | CodecFake+ (TASLP 2026) | Available |
| CoSG Eval | Eval | 866 | 931 | 17 CoSG systems | CodecFake+ (TASLP 2026) | Available |
| CoSG ExtEval | Ext. Eval | — | — | 40 CoSG systems | Proxy-to-Wild (ISCSLP 2026) | Coming soon |
- [2026.11] 🏆 Our follow-up paper, Mitigating Proxy-to-Wild Domain Gap in Deepfake Speech, received an Outstanding Paper award at ISCSLP 2026. It introduces the CoSG ExtEval evaluation set.
- [2026.05] CodecFake+ was published in IEEE Transactions on Audio, Speech and Language Processing. The earlier preprint title was "CodecFake+: A Large-Scale Neural Audio Codec-Based Deepfake Speech Dataset."
- [2025.12] 🏆 Our source tracing paper, Towards Generalized Source Tracing for Codec-Based Deepfake Speech, received a Best Student Paper Nomination (Top 2.5%) at IEEE ASRU 2025. [Award]
- [2025.10] The CoRS and CoSG sets and their label files were released on Hugging Face.
CodecFake+ is one part of our research line on codec-based deepfake speech. This table shows where it fits: the Dataset & Detection papers cover how CodecFake speech is detected and how well detectors generalize, and the Source Tracing papers use CodecFake+ to trace how a fake was generated, from defining the tasks to advancing the methods.
Dataset & Detection
| Paper | Position | Contribution | Builds on |
|---|---|---|---|
| Proxy-to-Wild (ISCSLP 2026) 🏆 Outstanding Paper |
In-the-wild generalization | Domain-Shift Feature Augmentation; CoSG ExtEval | Extends CoSG Eval: 40 unseen models, long-form audio |
| CodecFake+ (TASLP 2026) [arXiv] |
Large-scale benchmark | This dataset: CoRS, CoSG Eval, and the codec taxonomy | Extends CodecFake: 6 → 21 codec types, 3 → 17 CoSG systems, adds taxonomy |
| CodecFake (INTERSPEECH 2024) | Proof of concept | 6 codec types (15 codec models), 3 CoSG systems | — |
Source Tracing
| Paper | Position | Contribution | Builds on |
|---|---|---|---|
| SASTNet (ASRU 2025) 🏆 Best Student Paper Nomination (Top 2.5%) [Code] [Award] |
Advanced method | Semantic-Acoustic Source Tracing Network | Taxonomy-based source tracing tasks on CodecFake+ |
| Taxonomy-based source tracing (INTERSPEECH 2025) [ISCA] [Code] |
Task definition | Source tracing tasks via the codec taxonomy | CodecFake+ dataset and taxonomy |
Source tracing code and model checkpoints are available separately.
If you use CodecFake+ or CoSG ExtEval, please cite:
@article{chen2026codecfake+,
title = {{CodecFake+}: Codec-Based Resynthesized Data as a Proxy for Detecting {CodecFake} Speech},
author = {Chen, Xuanjun and Du, Jiawei and Wu, Haibin and Zhang, Lin and Lin, I-Ming and Chiu, I-Hsiang and Ren, Wenze and Tseng, Yuan and Tsao, Yu and Jang, Jyh-Shing Roger and Lee, Hung-yi},
journal = {IEEE Transactions on Audio, Speech and Language Processing},
volume = {34},
pages = {2929--2944},
year = {2026},
doi = {10.1109/TASLPRO.2026.3692291}
}
@inproceedings{chen2026mitigating,
title = {Mitigating Proxy-to-Wild Domain Gap in Deepfake Speech},
author = {Chen, Xuanjun and Wu, Yun-Shing and Lu, Wei-Chung and Lin, Claire and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
booktitle = {Proceedings of the International Symposium on Chinese Spoken Language Processing (ISCSLP)},
year = {2026}
}
@inproceedings{wu24p_interspeech,
title = {{CodecFake: Enhancing Anti-Spoofing Models Against Deepfake Audios from Codec-Based Speech Synthesis Systems}},
author = {Haibin Wu and Yuan Tseng and Hung-yi Lee},
year = {2024},
booktitle = {{INTERSPEECH 2024}},
pages = {1770--1774},
doi = {10.21437/Interspeech.2024-2093},
issn = {2958-1796},
}If you build on our source tracing work, please cite:
@inproceedings{chen2025towards,
title = {Towards Generalized Source Tracing for Codec-Based Deepfake Speech},
author = {Chen, Xuanjun and Lin, I-Ming and Zhang, Lin and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
booktitle = {2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
pages = {1--8},
year = {2025},
doi = {10.1109/ASRU65441.2025.11434771}
}
@inproceedings{chen2025codec,
title = {Codec-Based Deepfake Source Tracing via Neural Audio Codec Taxonomy},
author = {Chen, Xuanjun and Lin, I-Ming and Zhang, Lin and Du, Jiawei and Wu, Haibin and Lee, Hung-yi and Jang, Jyh-Shing Roger},
booktitle = {{INTERSPEECH 2025}},
pages = {1538--1542},
year = {2025},
doi = {10.21437/Interspeech.2025-1297}
}