List your company
ResearchFormerly WizWand

Audio Deepfake Detection on ASVspoof DF 2021 (EER)

1.43EERMamBo-4-Hydra-N1
0.5886.271511.95517.6385Sep 25, 2025Nov 10, 2025Dec 26, 2025Feb 10, 2026Mar 28, 2026May 13, 2026Jun 29, 2026
Updated 4mo ago

Evaluation results

MethodLinks
MamBo-4-Hydra-N1Params (M)=318.022026.011.43
Tran et al.Backbone=XLS-R, Backbone status=Fine-tuned, Layers used=25 (gated), Trainable params=318M§, Fusion=Gated2026.061.53
XLSR-Nes2NetXLoss Function=QAMO2025.091.6
XLSR-Conformer-TCMLoss Function=QAMO2025.091.63
MamBo-3-Hydra-N3Params (M)=319.372026.011.7
Fake-Mamba (L)Params (M)=319.722026.011.74
XLSR-MambaTraining Strategy=DPDA training + PCGrad2025.091.74
HyperPotterParams (M)=317.872026.021.78
W2V2-AASISTTraining Protocol=Protocol 2 (multi-dataset), Model Scale=317M2026.031.78
XLSR-Nes2NetXLoss Function=WCE, experimental_setting=Original2025.091.78
XLSR-Nes2Net-X2025.091.78
XLSR-Conformer-TCMTraining Strategy=DPDA training + PCGrad2025.091.81
mHuBERT-FinalTraining Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.031.83
MamBo-2-Hydra-N1Params (M)=316.682026.011.84
XLS-R+STCA+LMDC2026.021.87
XLSR-AASISTTraining Strategy=DPDA training2025.091.87
XLSR+MambaParams (M)=3192026.021.88
XLSR-MambaParams (M)=319.332026.011.88
MoLEXTraining Protocol=Protocol 2 (multi-dataset), Model Scale=376M2026.031.88
XLSR-MambaLoss Function=WCE2025.091.88
XLSR-Conformer-NACLLoss Function=WCE2025.091.88
XLSR-MambaTraining Strategy=Baseline2025.091.88
XLSR-Conformer-TCMLoss Function=OC-Softmax2025.091.89
XLSR+SLSParams (M)=3402026.021.91
XLSR-SLSParams (M)=341.492026.011.92
XLSR-SLS2025.091.92
XLS-R + HierConBackbone=XLS-R 300M, Head=HierCon2026.021.93
Wav2DF-TSL2025.091.95
MamBo-1-Mamba2-N2Params (M)=317.482026.012.01
XLSR-Conformer+TCMParams (M)=319.772026.012.06
XLSR-Conformer-TCMLoss Function=WCE, experimental_setting=Original2025.092.06
XLSR-Conformer-TCMTraining Strategy=Baseline2025.092.06
XLS-R + SLSBackbone=XLS-R, Head=SLS2026.022.09
XLSR-Conformer-TCMTraining Strategy=DPDA training2025.092.11
XLSR-AASISTTraining Strategy=DPDA training + PCGrad2025.092.13
XLSR+Conformer+TCMParams (M)=3192026.022.15
OC+ACS2026.022.19
ASC+OCLoss Function=WCE2025.092.19
OCKD2026.022.27
XLSR-ConformerParams (M)=319.742026.012.27
XLSR-Nes2NetXLoss Function=OC-Softmax2025.092.29
XLSR-MambaTraining Strategy=DPDA training2025.092.31
XLSR+BiCrossMambaParams (M)=318.212026.022.35
mHuBERT-Iter2Training Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.032.37
XLSR-Conformer-TCMLoss Function=WCE, experimental_setting=Reproduced2025.092.39
XLSR+MoE2026.022.54
XLSR-MoE2025.092.54
WavLM+MFA2026.022.56
WavLM+MFABackbone=WavLM, Head=MFA2026.022.56
XLSR+ConformerParams (M)=302.402026.022.58
XLSR-ConformerLoss Function=WCE2025.092.58
XLSR-Nes2NetXLoss Function=WCE, experimental_setting=Reproduced2025.092.76
XLSR+AASIST22026.022.77
Wav2Vec+AASIST2Backbone=Wav2Vec, Head=AASIST22026.022.77
FTDKD2026.022.82
Wav2Vec+AASISTBackbone=Wav2Vec, Head=AASIST2026.022.85
W2V2-TCMTraining Protocol=Protocol 2 (multi-dataset), Model Scale=319M2026.032.91
WavLM+AttMBackbone=WavLM, Head=AttM2026.023.19
Probing-Guided Layer SelectionBackbone=XLS-R, Backbone status=Frozen, Layers used=4, Trainable params=1.34M, Fusion=Concat2026.063.23
DF-Arena 500MTraining Protocol=Protocol 2 (multi-dataset), Model Scale=500M2026.033.3
XLSR-AASIST-SAM2025.093.44
WavLM-Base+Training Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.033.59
Wav2Vec+LinearBackbone=Wav2Vec, Head=Linear2026.023.65
HuBERT-BaseTraining Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.033.68
XLSR-AASISTTraining Strategy=Baseline2025.093.69
ResembleAI-2BTraining Protocol=Protocol 2 (multi-dataset), Model Scale=2000M2026.033.79
WavLM-BaseTraining Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.033.93
MLDG-LoRABackbone=W2V 2.0, Backbone status=LoRA, Layers used=All, Trainable params=3.59M, Fusion=Feature2026.063.99
Wav2Vec2+AASIST(Baseline)Params (M)=317.842026.024.08
WavLM+ASPBackbone=WavLM, Head=ASP2026.024.47
mHuBERT-Iter1Training Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.034.53
DF-Arena 100M-V1Training Protocol=Protocol 2 (multi-dataset), Model Scale=100M2026.035.5
Wav2Vec2+AASISTParams (M)=317.842026.026.63
Dual-Granularity Orthogonal DisentanglementFine-tuning status=without fine-tuning, Variant=Full Model2026.067.88
WavLM-MLPPre-training (SSL-based)=true, Model Parameters=>300M, Fine-tuning status=without fine-tuning2026.067.95
Dual-Granularity Orthogonal Disentanglement (Cosine Only)Fine-tuning status=without fine-tuning, Variant=Cosine Only2026.068.23
DG-AggFine-tuning status=without fine-tuning2026.068.26
Wav2Vec2-AASISTPre-training (SSL-based)=true, Model Parameters=>300M, Fine-tuning status=without fine-tuning2026.068.54
GRL BaselineFine-tuning status=without fine-tuning2026.068.91
Dual-Granularity Orthogonal Disentanglement (No Disentangle)Fine-tuning status=without fine-tuning, Variant=No Disentangle2026.069.05
Res-TSSDNetFine-tuning status=without fine-tuning2026.069.81
Dual-Granularity Orthogonal Disentanglement (Single Branch)Fine-tuning status=without fine-tuning, Variant=Single Branch2026.0610.17
RawGAT-STFine-tuning status=without fine-tuning2026.0610.75
AASISTFine-tuning status=without fine-tuning2026.0612.83
BiCrossMamba-STParams (M)=0.5162026.0114.77
RawBMambaParams (M)=0.7192026.0115.85
RawMambaParams (M)=0.7192026.0122.48