Preventing converted speech spoofing attacks in speaker verification

Voice conversion (VC) techniques, which modify a speaker´s voice to sound like another´s, present a threat to automatic speaker verification (SV) systems. In this paper, we evaluate the vulnerability of a state-of-the-art SV system against a converted speech spoofing attack. To overcome the spoofing attack, we implement state-of-the-art converted speech detectors based on short- and long-term features. We propose a new converted speech detector using a compact feature representation and a discriminative modeling approach. We experiment pairing converted speech detectors based on short- and long-term features to improve converted speech detection. The results indicate that the proposed converted speech detector pair outperforms state-of-the-art ones, achieving a detection accuracy of 97.9% for natural utterances and 98.0% for converted utterances. We include the anti-spoofing mechanism in our SV system as a post-processing module for accepted trials and reevaluate its performance, comparing it with the performance of an ideal system. Our results show that the SV system´s performance returns to acceptable values, with less than 1.6% equal error rate (EER) change.