When Automatic Voice Disguise Meets Automatic Speaker Verification

The technique of transforming voices in order to hide the real identity a speaker is called voice disguise, among which automatic disguise (AVD) by modifying spectral and temporal characteristics with miscellaneous algorithms are easily conducted softwares accessible public. AVD has posed great threat both human listening verification (ASV). In this paper, we have found that ASV not only victim but could be tool beat some simple types AVD. Firstly, three AVD, pitch scaling, vocal tract length normalization (VTLN) conversion (VC), introduced as representative methods. State-of-the-art methods subsequently utilized objectively evaluate impact on equal error rates (EER). Moreover, an approach restore disguised its original version proposed minimizing function scores w.r.t. restoration parameters. Experiments then from Voxceleb, dataset recorded real-world noisy scenario. results shown that, for obtains EER around 7% comparing 30% recently baseline using ratio fundamental frequencies. generalizes well nonlinear frequency warping VTLN reducing 34.3% 18.5%. However, it difficult source speakers VC our approach, where more complex forms functions or other paralinguistic cues might necessary transform VC. Finally, contrastive visualization features without illustrate role intuitive way.