The experiment uses on spectrogram-based features with logistic regression classifier
dev_eer | 0.0636682 |
---|---|
dev_eer_threshold | 0.201115 |
dev_far | 0.0636599 |
dev_frr | 0.0636764 |
test_far | 0.0606187 |
test_frr | 0.0563128 |
test_hter | 0.0584657 |
dev_negatives | 38580 |
dev_positives | 4994 |
test_negatives | 43320 |
test_positives | 5576 |
dev_scores | |
test_scores | |
dev_roc | |
test_roc |
A simple baseline of speaker spoofing attacks detection system.
This algorithm relies on the Bob library.
The baseline system relies on: * VAD detection based on 4Hz modulation * Audio spectrogram * Feature vectors as ratios between average values of the specified number of subbands of the spectrogram * Logistic regression classifier which is trained on the training set and computes score on the development set.
As the results, EER value is computed and Detection Error Tradeoff (DET) curve is plotted.