LaST-Net: Speech Anti-Spoofing Arena wrapper
Packaged for the Speech Anti-Spoofing Arena from the authors' code.
No paper: the model sits in the Arena's unpublished tier (shown, not ranked).
LaST-Net (Length-Aware Layer and Scale-Adaptive Temporal Network; Jun Xue, 2026): a fine-tuned fairseq XLS-R 300M front-end whose 24 transformer-layer outputs are fused by a length- and content-conditioned layer router, projected to 192-d, processed at three temporal scales (1x, 1/2, 1/4) by a length-selective kernel block and dilated depthwise residual blocks with two length-prior residual scale-exchange stages, attentive statistics pooling per scale and a 2-way MLP head. The model also takes the clip's available duration (length features). Trained on ASVspoof 2019 LA train (dev for checkpoint selection) with RawBoost-4 paired prediction consistency. Authors' released epoch-52 checkpoint. Input exactly as the authors' infer.py at --seconds 4: 16 kHz, no normalisation, first 64000 samples (4 s), shorter clips repeated (np.tile) and cropped, available seconds = min(len, 64000)/16000 passed to the model. score = log_softmax(logits)[:, 1] (the authors' bonafide_log_score). FP32 (TF32 off).
- Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the
meta.yamlnotes. - Params: 316.96 M. Trained on ASVspoof 2019 LA train (dev for selection); the Arena's ASVspoof sets are eval partitions.
- 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; FP32 on the RTX 5060 Ti), 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/.
- Downloads last month
- 188