RawNetLite: Speech Anti-Spoofing Arena wrapper

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

RawNetLite (Di Pierno et al., IJCNN VERIMEDIA 2025): a lightweight end-to-end raw-waveform detector -- Conv1D(1->64) + BatchNorm + ReLU, three 1-D residual blocks (64 ch), adaptive average pooling to 64 steps, a bidirectional GRU (128 per direction, last step), Linear(256,64) -> Linear(64,1). Released checkpoint triple_cross_domain_focal_rawnet_lite.pt: trained with focal loss on a balanced mix of FakeOrReal, "AVSpoof2021" (the authors' name for ASVspoof 2021) and CodecFake, no augmentation. Input as the authors' audio_preprocessor.py: 16 kHz mono, peak-normalised (divide by max |x| of the whole clip), first 48000 samples (3 s, within the Arena 4 s cap), right zero-pad. score = -logit(fake) (the pre-sigmoid fc2 output negated; the source's sigmoid(logit) is P(fake)). FP32 (TF32 off).

Arena results

arena tier arena rank

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/.

EER% 15.08 on ASVspoof2019_LA EER% 18.14 on ASVspoof2021_LA EER% 20.19 on ASVspoof2021_DF EER% 45.69 on ASVspoof5 EER% 40.51 on InTheWild EER% 44.11 on ADD22_eval_31 EER% 46.63 on ADD2023_track12_test_r1 EER% 34.22 on CFAD EER% 42.55 on CD-ADD EER% 37.46 on CVoiceFake_small EER% 50.48 on DeepVoice EER% 51.27 on LibriSeVoc EER% 42.04 on SONAR EER% 42.98 on ArAD EER% 35.72 on DECRO EER% 34.19 on J-SPAW_LA EER% 48.1 on ODSS EER% 33.38 on HABLA EER% 49.93 on DFADD EER% 43.63 on PyAra EER% 48.2 on XMAD EER% 61.18 on EmoFake_test SRR COMPLEMENT 5.66 on EmoSpoofTTS SRR COMPLEMENT 10.48 on LRLspoof

Citation

@inproceedings{dipierno2025rawnetlite,
  title     = {End-to-end Audio Deepfake Detection from RAW Waveforms: a RawNet-Based Approach with Cross-Dataset Evaluation},
  author    = {Andrea Di Pierno and Luca Guarnera and Dario Allegra and Sebastiano Battiato},
  booktitle = {International Joint Conference on Neural Networks (IJCNN) - VERIMEDIA Workshop},
  year      = {2025}
}
Downloads last month
185
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Paper for SpeechAntiSpoofingBenchmarks/RawNetLite