XLSR-Mamba-LA — Speech Anti-Spoofing Arena wrapper

Xiao & Das, IEEE SPL — XLSR-Mamba: a dual-column bidirectional Mamba on XLS-R. Two checkpoints released (LA- and DF-tuned RawBoost); not to be confused with XLSR-MamBo (arXiv 2601.02944).

Packaged for the Speech Anti-Spoofing Arena from the authors' code and paper.

XLSR-Mamba (LA checkpoint): a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end (last layer) projected to 144-d (Linear + BatchNorm + SELU), followed by a dual-column bidirectional Mamba (v1 selective SSM) back-end: 6 forward + 6 backward Mamba blocks with fused add+RMSNorm, per-direction attention pooling, concatenation, linear fusion and a 2-way head. Official authors' LA-track checkpoint (trained on ASVspoof2019 LA train with RawBoost algo 5). Not to be confused with XLSR-MamBo (arXiv 2601.02944), a different model.

• Checkpoint (commit-pinned): model.safetensors
• Params: 319.3275 M. Trained on ASVspoof2019 LA train only.
• Score: logits[:, 1] (higher = bona fide). Input: first 66,800 samples @16 kHz, shorter clips repeat-padded; FP32. J-SPAW_LA was scored with the PyTorch wrapper; the other datasets with the TensorRT drop-in class (FP32), which matched PyTorch on the full J-SPAW_LA set.

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{xiao2024xlsr,
  title   = {{XLSR-Mamba}: A Dual-Column Bidirectional State Space Model for Spoofing Attack Detection},
  author  = {Xiao, Yang and Das, Rohan Kumar},
  journal = {IEEE Signal Processing Letters},
  note    = {arXiv:2411.10027},
  year    = {2025}
}