UC Davis

Speakers vary in how they produce language, and this variation can be stylistic, socially motivated, or based on individual differences. In human language, all communication has social meaning (Giles & Ogay, 2013; Eckert, 2012). But, it is unknown how the same variation would be perceived in “intelligent” computers like voice AI assistants and chatbots. Previous research has shown that when faced with increasing cues of humanity, users apply gender stereotypes, show politeness, linguistically align with, and anthropomorphize social computers (Cohn et al., 2023; Nass, et al., 1997; Sundar & Kim, 2012). Yet, it is unknown whether users evaluate social computer behavior as socially meaningful. Because of the importance of sociolinguistic variation in the generation of social meaning, the use of social-indexical variation by computer voices is an apt way to investigate whether users evaluate computers as social actors.In three experiments, this dissertation research investigated listener evaluations of the sociolinguistic (ING) variable when used by human and computer talkers. Experiment 1 investigated how listeners perceive (ING) variation in human voices compared to computer ones. I found that listeners transfer the social meaning of the standard -ing variant to their evaluations of appropriateness for broadcasting and social status, even when the voice was a computer. However, (ING) variation was more influential on listeners’ evaluations when it was used by human talkers, and human talkers were evaluated more positively regardless of (ING) use. Experiment 2 tested how college-age students perceive variation in synthetic voices in two different speaking styles. In Experiment 2, which used only synthetic voice stimuli, I found that participants’ perception of a voice as a human was primarily influential on their social evaluations, but interactions between speaking style, participant gender, and (ING) variation were also impactful. Experiment 3 used the same design as Experiment 2 to test how older adults evaluate synthetic voices. Older adults evaluated voices they perceived as computers especially negatively, particularly when the voices used nonstandard -in’. I found that older adults’ personal voice AI use habits and the talker’s speaking style were influential on their evaluations, but I did not find large gender effects like in Experiment 2.Across studies and across voice types, listener evaluations interact with listener and speaker indexical traits like age, gender, and speaking style, similar to previous research on (ING) in human speech. Belief that a voice was a computer was one of many interrelated social perceptions that influenced listeners’ rating behavior, suggesting that computer voices have their own “social class” in the mind of listeners. In other words, synthetic voices that use sociolinguistic variation are evaluated with the social meaning of (ING) from human communication but mediated by perception of human likeness and listeners’ personal traits.Looking forward, the ever-increasing sophistication of synthetic voices and social computer design will have complex implications for listener perception of computer voices. Future linguistically informed research in the fast-advancing field of human-computer interaction is essential to understand how people socially relate to nonhuman interactants.