SSL-Mamba: Speech Anti-Spoofing Arena wrapper

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

SSL-Mamba (Tran et al., Interspeech 2025): a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end whose 25 hidden states are each projected to 128-d (Linear + SiLU) and summed (self-gated layer aggregation), a multi-kernel gated convolution (MultiConv cgMLP, kernels 7/15/23/31) fusion module, 12 unidirectional Mamba (v1 selective SSM) residual blocks with LayerNorm, multi-head attention pooling (4 heads, mean + std) and a 2-way head. Official authors' checkpoint (interspeech2025.pt), trained on ASVspoof2019 LA train. Not to be confused with XLSR-Mamba (Xiao & Das), a different model.

• Checkpoint (commit-pinned): model.safetensors
• Params: 317.5213 M. Trained on ASVspoof 2019 LA train.
• Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; the 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.

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{tran25b_interspeech,
  title     = {{Leveraging SSL Speech Features and Mamba for Enhanced DeepFake Detection}},
  author    = {Tran, Hoan My and Lolive, Damien and Guennec, David and Sini, Aghilas and Delhay, Arnaud and Marteau, Pierre-Fran{\c{c}}ois},
  booktitle = {Interspeech 2025},
  pages     = {5323--5327},
  year      = {2025},
  doi       = {10.21437/Interspeech.2025-1703}
}