The authors propose a linguistic-invariant spoofing detection framework that addresses poor generalization in out-of-domain settings caused by reliance on linguistic cues. The method uses a teacher-student adversarial learning approach where a linguistic-aware teacher guides the student detector via gradient reversal to minimize linguistic information.

  • A Variational Information Bottleneck is incorporated to prevent the inadvertent removal of non-linguistic cues while suppressing principal linguistic ones.
  • The framework was evaluated across nine DF Arena datasets.
  • The method achieves up to a 36.2% relative reduction in Equal Error Rate (EER) compared to the baseline.

This approach improves robustness to cross-data scenarios by decoupling spoofing detection from specific linguistic content.