NII-SLT26-XLSR-FC: Speech Anti-Spoofing Arena wrapper
Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.
XLS-R + FC from Ge, Wang and Yamagishi (IEEE SLT 2026): a fully fine-tuned wav2vec 2.0 XLS-R 300M encoder (transformers Wav2Vec2Model), mean pooling of the last hidden state over all frames, and a Linear(1024, 2) head [real, fake]. Trained on CompSpoof V2 (ESDD2) train with the "any component fake = fake" labels, the unmixed "original" class excluded (protocol AnyFakeNoOriginal), seed 2, best-validation-loss checkpoint. Authors' released checkpoint. Input as the authors' collator: 16 kHz waveform, per-clip zero-mean unit-variance normalisation (Wav2Vec2FeatureExtractor), applied after the Arena 4 s cap (first 64600 samples, shorter clips repeat-padded; the authors score whole clips). score = logit[real] - logit[fake] (the authors' Score_real - Score_fake on softmax probabilities is a monotone function of it). FP32 (the authors' eval precision; TF32 off).
- Weights: re-hosted here under the upstream license (nii_slt26_xlsr_fc.safetensors, commit-pinned); upstream: https://github.com/nii-yamagishilab/AntiDeepfake-AudioLLM/tree/2d445b525bb52518448b353fe1fee51d5f815af9
- Params: 315.44 M. Trained on CompSpoof V2 train: bona fide CommonVoice, LibriTTS and english-conversation-corpus speech mixed with environmental sound; spoof speech incl. ASVspoof 5, SONAR, CD-ADD, LibriSeVoc and MLAAD sources. The Arena marks ASVspoof5, SONAR, CD-ADD and LibriSeVoc as n/a* (training overlap). Clean studio bona fide speech is out of domain for these checkpoints; several cells are near-random or inverted and are kept (maintainer's choice).
- Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; most other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; PyTorch on the RTX 5060 Ti (no engine built there)), which matched PyTorch on the full J-SPAW_LA set (|dEER| <= 0.25, Spearman >= 0.999); the RTX 4060 Ti cards ran PyTorch.
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/.
Citation
@inproceedings{ge2026what,
title = {What Do Deepfake Audio Detectors Hear? From Self-Supervised Models to Audio LLMs},
author = {Ge, Wanying and Wang, Xin and Yamagishi, Junichi},
booktitle = {IEEE Spoken Language Technology Workshop (SLT)},
year = {2026}
}
- Downloads last month
- 180