Audio Deepfake Detection on ASVspoof 2021
1.92EERXLS-R + SLS
Evaluation results
| Method | Links | ||
|---|---|---|---|
| XLS-R + SLSArchitecture / Strategy=Self-Sup. XLS-R + Layer Selection2025.04 | 1.92 | — | |
| ACS One-ClassArchitecture / Strategy=Adaptive Centroid Shift (1-class)2025.04 | 2.19 | — | |
| WavLM + MFAArchitecture / Strategy=WavLM + Multi-Fusion Attention2025.04 | 2.56 | — | |
| Tak et al.Architecture / Strategy=XLS-R + RawNet2 + AASIST Backend2025.04 | 2.85 | — | |
| XLSR-SLSTraining Setting=ASVspoof 2019 LA2026.04 | 3.04 | — | |
| W2V-AASISTCategory=Finetuned frontend, #Param (million)=3172024.07 | 3.6 | 0.707 | |
| ProSDDTraining Setting=ASVspoof 2019 LA2026.04 | 3.87 | — | |
| SLIMCategory=Finetuned frontend, #Param (million)=2532024.07 | 4.4 | 0.651 | |
| W2V-ASPCategory=Finetuned frontend, #Param (million)=3172024.07 | 4.5 | 0.646 | |
| RawNet2Training Setting=ASVspoof 2019 LA2026.04 | 8.08 | — | |
| W2V-LCNNCategory=Frozen frontend, #Param (million)=N/A2024.07 | 8.1 | — | |
| AASISTTraining Setting=ASVspoof 2019 LA2026.04 | 8.15 | — | |
| FullCategory=SLIM variants (ours), #Param (million)=112024.07 | 8.3 | 0.451 | |
| Enc_styCategory=SLIM variants (ours), #Param (million)=92024.07 | 8.6 | 0.438 | |
| SSL-FusionCategory=Frozen frontend, #Param (million)=102024.07 | 8.9 | 0.419 | |
| WLM-ASPCategory=Frozen frontend, #Param (million)=92024.07 | 9 | 0.426 | |
| Enc_style+lingCategory=SLIM variants (ours), #Param (million)=102024.07 | 9 | 0.429 | |
| Enc_lingCategory=SLIM variants (ours), #Param (million)=92024.07 | 9.3 | 0.407 | |
| W2V-LLGFCategory=Frozen frontend, #Param (million)=102024.07 | 9.4 | 0.402 | |
| W2V+HUBCategory=Frozen frontend, #Param (million)=92024.07 | 14.2 | 0.31 | |
| HUB-ASPCategory=Frozen frontend, #Param (million)=92024.07 | 15.4 | 0.289 | |
| RawNetLite (Cross-domain)Architecture / Strategy=Raw waveform + GRU + AVS mix2025.04 | 16.6 | — | |
| WLM+HUBCategory=Frozen frontend, #Param (million)=92024.07 | 16.7 | 0.269 | |
| RawNetLite (Cross + Focal)Architecture / Strategy=+ Focal Loss2025.04 | 17.3 | — | |
| ProSDDTraining Setting=ASVspoof 20242026.04 | 18.08 | — | |
| ProSDDTraining Setting=Trained on ASVspoof 20242026.04 | 18.08 | — | |
| RawNetLite (Triple + Augmented)Architecture / Strategy=+ Augmentations2025.04 | 18.1 | — | |
| W2V-ASPCategory=Frozen frontend, #Param (million)=92024.07 | 19.6 | 0.233 | |
| RawNetLite (Triple + Focal)Architecture / Strategy=+ CodecFake2025.04 | 20 | — | |
| DependencyCategory=SLIM variants (ours), #Param (million)=92024.07 | 20.5 | 0.234 | |
| RawNet2Category=Baseline, #Param (million)=42024.07 | 22.3 | 0.213 | |
| W2V+WLMCategory=Frozen frontend, #Param (million)=92024.07 | 22.5 | 0.203 | |
| AASISTTraining Setting=ASVspoof 20242026.04 | 22.74 | — | |
| AASISTTraining Setting=Trained on ASVspoof 20242026.04 | 22.74 | — | |
| LCNNCategory=Baseline, #Param (million)=42024.07 | 25.5 | 0.197 | |
| RawNet2Training Setting=ASVspoof 20242026.04 | 25.59 | — | |
| RawNet2Training Setting=Trained on ASVspoof 20242026.04 | 25.59 | — | |
| XLSR-SLSTraining Setting=ASVspoof 20242026.04 | 26.54 | — | |
| XLSR-SLSTraining Setting=Trained on ASVspoof 20242026.04 | 26.54 | — |