MMS-1B-AntiDeepfake-NDA

Bit-exact copy of nii-yamagishilab/mms-1b-anti-deepfake-nda (the AntiDeepfake collection, nii-yamagishilab), hosted for the Speech Anti-Spoofing Arena. Trained without data augmentation (NDA).

• Paper: Post-training for Deepfake Speech Detection, arXiv:2506.21090
• Original code: https://github.com/nii-yamagishilab/AntiDeepfake
• Backbone: MMS-1B; head = AdaptiveAvgPool1d + Linear(D, 2). Score = real logit.
• License: CC BY-NC-SA 4.0 (non-commercial) — inherited from the source.

⚠️ Evaluation restriction

It is STRICTLY PROHIBITED to evaluate this model on its training / in-domain corpora: ASVspoof2019 LA, ASVspoof2021 LA/DF, ASVspoof5, CFAD, DECRO, DFADD, "Diffuse or Confuse", DiffSSD, DSD, HABLA, MLAAD, SpoofCeleb, CVoiceFake, any LibriTTS-based set (incl. SONAR, CD-ADD, LibriSeVoc), any VoxCeleb-based set, WaveFake, any FLEURS-based set, CodecFake (any variant), any AIShell3-based set, and any MLS-based set. On the Arena it is evaluated only on ArAD and In-the-Wild.

All credit to the original authors (Ge, Wang, et al., NII).

Arena results