XLSR-MamBo: Scaling the Hybrid Mamba-Attention Backbone for Audio Deepfake Detection
Paper β’ 2601.02944 β’ Published
XLSR-MamBo (MamBo-3, Hydra-N3) β Speech Anti-Spoofing Arena wrapper
Official checkpoint of XLSR-MamBo: Scaling the Hybrid Mamba-Attention Backbone for Audio Deepfake Detection (Ng et al., ACL 2026 Findings β paper, arXiv 2601.02944, code, MIT), packaged for the Speech Anti-Spoofing Arena.
- Checkpoint:
mambo3_da0_best0.pt= authors' MamBo-3 (Hydra-N3), DA0 (no RawBoost),best_0.pt(lowest ASVspoof2019 LA validation loss; =model_best.pt). 319.37 M params. Trained on ASVspoof2019 LA. - Input: first 66 800 samples @16 kHz, shorter clips repeat-padded. Score =
logit[bonafide] - logit[spoof](higher = bona fide). xlsr_mambo.py(PyTorch wrapper),_net.py(network),mamba_ssm/(vendored Triton subset of mamba_ssm 2.2.6.post3, Apache-2.0),_hydra/(Hydra mixer from the source repo),trt_xlsr_mambo.py(TensorRT FP32 drop-inXLSRMamBoTRT).- Deviations from the source: per-utterance waveform layer-norm (torchaudio's wrapper normalises jointly over the batch); FP32 instead of bf16 autocast; SSD scan + RMSNorm in plain PyTorch (same math as the Triton kernels, max |Ξscore| 2e-4).
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
- -
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support