Dialect imitations in speaker recognition

Abstract

Voice imitation and disguise are possible threats to the performance of a speaker recognition system and to the accuracy of earwitness descriptions. One common disguise is the modification of the own dialect or accent. In this paper, this kind of disguising is explored, using recordings from a well-known actor with considerable experience of dialect and accent imitation. In order to see how successful his dialect imitations are and how the process of speaker discrimination is influenced by accent disguise, two sets human perception tests were constructed. One set focused on American and British English dialects, and one set on American and London English accents and Spanish-accented English. Each set consisted of three parts: a same-different speaker test, a same-different accent test, and a select the accent from a closed-set of options test. The results show that Johnny Depp is successful without his visual props and demonstrate a high correlation between the quality of the accent imitations and the failure of the human listeners to recognize that the voices come from the same speaker. The third parts of the experimental sets suggest the importance of familiarity with the accent that feeds into parts one and two. Spanish listeners, for example, are less accepting of the Spanish-accented English than non- Spanish speakers. Finally, the same speech segments used in the perception test were used in an automatic speaker recognition experiment in order to compare the results and to check the robustness of the system in front of the voice changes. The results showed, once again, a low correlation between human and automatic speaker recognition.