Comparative Analysis of Voice Mimicry Attacks by High- and Low-Skilled Imitators on Speaker Verification Systems

Given that "voice mimicry" constitutes a fundamental attack against speaker verification systems, it is essential to comprehend the impact of these attacks on the system and their characteristics. This study evaluates and compares the effects of voice mimicry attacks by high- and low-skilled imitators on speaker verification systems under impartial conditions. Two types of speaker verification systems are used in the analysis: a classical system based on GMM-UBM method and a deep-learning-based system using x-vector. Results of the comparative analysis, using a dataset originally constructed, indicate that the attack of the high-skilled imitator has higher spoofing ability against both systems than the low-skilled imitators. Furthermore, the analysis of the verification scores obtained from the x-vector-based speaker verification system revealed differences in the characteristics of high- and low-skilled imitators in voice mimicry attacks.