SONAR-Full

Mirror of the official SONAR-Full checkpoint (sonar_full_xlsr_aasist_eer6.pth) from idonithid/SONAR-Audio-DF-Detection, evaluated on the Speech Anti-Spoofing Arena.

Paper: Hidekel, Lifshitz, Cohen, Raviv. SONAR: Spectral-Contrastive Audio Residuals for Generalizable Deepfake Detection. ICML 2026 (PMLR 306:42998-43012). PMLR · arXiv 2511.21325

Dual XLS-R 300M encoders (raw waveform + constrained-SRM high-frequency residual), fused by cross-attention, with an AASIST graph-attention back-end. ~641.5M parameters.

Benchmark wrapper

• sonar_full.py: SONARFull(AntiSpoofingModel) for speech-spoof-bench
• _net.py: architecture vendored from the source repo (sonar/guided_model.py + sonar/srm_filters.py)
• Input: first 64,600 samples at 16 kHz (about 4 s), zero-padded. Score = logit of class 1 (real); higher = more bona fide.
• The XLS-R 300M base (xlsr2_300m.pt, fairseq) is needed only to build the architecture; every weight is overwritten by this checkpoint.

Arena results

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.

License

Weights: CC-BY-NC-4.0 (derivative of XLS-R). Source code: MIT.