RawBMamba

RawBMamba: End-to-End Bidirectional State Space Model for Audio Deepfake Detection (Chen et al., Interspeech 2024). A sinc/Res front-end feeds a bidirectional Mamba (selective state-space) backbone with attention pooling and an angular-margin (A-softmax) head, operating directly on the raw waveform. The model returns a score where higher = more bona fide.

• Original code: https://github.com/cyjie429/RawBMamba
• Paper: https://arxiv.org/abs/2406.06086
• Parameters: 720,180 (0.72 M)
• Checkpoint: rawbmamba_best.pt (the authors' released pretrained model)

This repo is self-contained for inference: the vendored mamba_ssm backbone and network are in mamba_ssm/ + _net.py, and the exact wrapper used to produce the Arena scores is in rawbmamba.py.

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/.

Architecture & inference notes

• Input: raw float32 mono @ 16 kHz, deterministic first-64600-sample (≈4 s) window (tile-repeat if shorter, no random crop). FP32.
• Score = cos_theta[:, 1] - cos_theta[:, 0] (the AngleLinear head output the repo's evaluation uses).
• Selective-scan path: the vendored mamba_ssm requires the causal_conv1d and selective_scan_cuda CUDA extensions for its fused fast path. We make those optional and run the pure-PyTorch reference selective-scan (use_fast_path off, fused_add_norm off) — numerically faithful to the fused kernels and, importantly, traceable for ONNX/TensorRT export.

ONNX / TensorRT

ONNX export + drop-in TensorRT inference class in trt_rawbmamba.py (RawBMambaTRT). Export required making the CUDA kernels optional, replacing the triton RMSNorm with a pure-PyTorch equivalent, and precomputing the AngleLinear renormalized weight (avoids the unsupported aten::renorm). The ONNX is uploaded here; the GPU-arch-specific engine is built locally (sidecar trt_rawbmamba.json).

Credits

All model design and weights are the work of the original RawBMamba authors (https://github.com/cyjie429/RawBMamba). This repository only packages their released checkpoint for the Speech Anti-Spoofing Arena.

@inproceedings{chen2024rawbmamba,
  title     = {RawBMamba: End-to-End Bidirectional State Space Model for Audio Deepfake Detection},
  author    = {Chen, Yujie and Yi, Jiangyan and Tao, Jianhua and others},
  booktitle = {Interspeech 2024},
  year      = {2024}
}