DF Arena 100M V0 - Antispoofing Model
We are excited to release DF Arena 100M V0 Antispoofing model trained on large scale speech antispoofing datasets. Check out the release on DF Arena leaderboard
Live Demo
Stay tuned for demo of the model to be released soon allowing users to test their samples for machine generated speech!
Evaluation
| Dataset | EER (%) | F1-score | Accuracy (%) |
|---|---|---|---|
| in_the_wild | 4.29 | 0.97 | 95.71 |
| asvspoof_2019 | 4.16 | 0.83 | 95.83 |
| asvspoof_2021_la | 10.06 | 0.64 | 89.94 |
| asvspoof_2021_df | 8.83 | 0.37 | 91.17 |
| asvspoof_2024 | 21.28 | 0.60 | 78.72 |
| fake_or_real | 6.14 | 0.94 | 93.84 |
| codecfake | 10.39 | 0.84 | 89.61 |
| add_2022_track_1 | 31.96 | 0.55 | 68.05 |
| add_2022_track_3 | 6.69 | 0.82 | 93.31 |
| add_2023_round_1 | 12.69 | 0.91 | 87.31 |
| add_2023_round_2 | 19.01 | 0.86 | 80.99 |
| dfadd | 0.13 | 0.99 | 99.89 |
| librisevoc | 0.42 | 0.99 | 99.58 |
| sonar | 5.73 | 0.95 | 94.30 |
| Average | 10.12 | 0.80 | 89.07 |
| Pooled | 15.93 | 0.70 | 84.06 |
License and Usage
The model uses a closed license. Kindly contact ajinkya.kulkarni@idiap.ch for access to the model.