Search papers, labs, and topics across Lattice.
Brno University of Technology
7
0
6
Human listeners misclassify genuine audio as fake 77% of the time, revealing a critical vulnerability in our ability to discern deepfake speech.
Deepfake speech detection may achieve sub-1% error rates in controlled settings, but real-world performance falters dramatically due to unforeseen challenges.
Global anchoring outperforms pairwise verification in synthetic speech source tracing, revealing hidden pitfalls in the latter's approach to metric learning.
Achieving a 20.7% reduction in compute while enhancing anti-spoofing robustness could redefine deployment strategies in self-supervised learning.
Fairness assessments of deepfake speech detectors are fundamentally flawed due to a lack of demographic metadata in most datasets, limiting meaningful subgroup analysis.
Deepfake speech detectors reveal surprising reliance on distinct audio cues, challenging assumptions about their decision-making processes.
Training with speaker references might seem essential, but RAT shows that models can excel in deepfake detection even when those references are absent during inference.
