Automatic versus human speaker verification: The case of voice mimicry

摘要: In this work, we compare the performance of three modern speaker verification systems and non-expert human listeners in presence voice mimicry. Our goal is to gain insights on how vulnerable are mimicry attack it listeners. We study both traditional Gaussian mixture model-universal background model (GMM-UBM) an i-vector based classifier with cosine scoring probabilistic linear discriminant analysis (PLDA) scoring. For studied material Finnish language, decreased lightly equal error rate (EER) for GMM-UBM from 10.83 10.31, while EER increased 6.80 13.76 4.36 7.38. The listening panel shows that imitated speech increases difficulty task. It even more difficult recognize a person who intentionally concealing his or her identity. Impersonator A, average listener made 8 errors 34 trials automatic had 6 same set. B 7 28 trials, 9 errors. A statistical was also conducted. found out statistically significant association, p ¼ 0:00019 R 2 0:59, between accuracy self reported factors only when familiar voices were present test.