WST-Graph-Q82: Speech Anti-Spoofing Arena wrapper
Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.
WST-Graph (Ng et al., 2026), configuration Q82: a fixed, parameter-free Kymatio wavelet scattering transform (J=8, Q=(8, 2), first + second order, oversampling 1) with log-modulation path normalisation, arranged as a modulation-carrier grid (OrderGrid), length-aware attentive pooling to 64 relative-time bins (ALAP), a channel adapter + depthwise-separable latent blocks and an AASIST graph back-end (GraphAASIST). Trained on ASVspoof 2019 LA train (focal loss, 12 epochs), training seed 9 (the paper names no headline seed; seed 9 is the model card's example). Authors' traced TorchScript program, run as released. Input as the model card: 16 kHz, no normalisation, first 64000 samples (4 s), right zero-pad, input_length = retained sample count. score = logit[bona fide] - logit[spoof] (logits[:,0]-logits[:,1]). FP32 (TF32 off).
- Weights: re-hosted here under the upstream license (wst_graph_q82_seed9_batch32_cuda.ts, commit-pinned); upstream: https://huggingface.co/kwokho1/wst-graph/tree/080000c1685451c7759f43d2ae7d0f08477cfdf8
- Params: 0.12 M. Trained on ASVspoof 2019 LA train (dev for selection). A weak cross-domain model: near-random cells are kept (maintainer's choice).
- Score: higher = bona fide. All datasets were scored with the PyTorch wrapper (7B audio LLM in bf16; no TensorRT).
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
@article{ng2026wstgraph,
title = {WST-Graph: Topology-Preserving Wavelet Scattering Front-End for Speech Deepfake Detection},
author = {Ng, Kwok-Ho and Song, Tingting and Feng, Bingwen and Xia, Zhihua},
journal = {arXiv preprint arXiv:2609.29372},
year = {2026}
}
- Downloads last month
- 208