Allison Jones ^* Georgia Zellou

• University of California, Davis, Davis, United States

The current study had two goals: First, we aimed to conduct a conceptual replication and extension of a classic study by Nass and colleagues (1997) who found that participants display voice-gender bias when completing a tutoring session with a computer. In the present study, we used a more modern paradigm (i.e., app-based tutoring) and commercially-available TTS voices. Second, we asked whether participants provided different social evaluations of non-native-accented and native-accented American English-speaking machines. In the present study, 85 American participants completed a tutoring session with a system designed to look like a device application (we called it a "TutorBot"). Participants were presented with facts related to two topics: 'love and relationships' and 'computers and technology'. Tutoring was provided either by a female or male TTS voice. Participants heard either native-English accented voices or non-native-English accented (here, Castilian Spanish-accented) voices. Overall, we find no effect of voice gender on any of the dependent measures: listeners recalled facts and rated female and male voices equivalently across topics and conditions. Yet, participants rated non-native accented TTS voices as less competent, less knowledgeable, and less helpful after completing the tutoring session. Finally, when participants were tutored on facts related to 'love and relationships', they showed better accuracy at recall and provided higher ratings for app competency, likeability, and helpfulness (and knowledgeable, but only for native-accented voices). These results are relevant for theoretical understandings of human-computer interaction, particularly the extent to which human-based social biases are transferred to machines, as well as for applications to voice-AI system design and usage.