Poster #79 - Hey Sila! How children speak with their voice assistant depends on what they think about her

• In Event: Poster Session 05

Abstract

Children encounter digital media at an increasingly early age. The growing relevance of AI-based tools in their lives provides children with the opportunity to communicate with a novel type of interaction partner: non-human interlocutors. However, we know far too little about how children communication with these devices. In particular, how do they talk to digital interaction partners and do they differentiate between humans and AI-based interaction partners in digital communication? Also, do children’s assumptions about their digital communication partners influence the way they interact with them?
To address these aspects, we conducted a treasure hunt game with 50 5–6-year-olds who interacted with a digital assistant, called Sila, who was either introduced as a human or a voice assistant (between-subjects; Figure 1 for the study design). In both groups, the digital agent remotely guided children through the hunt and assisted in finding boxes which provided new keys with labels for subsequent boxes, eventually leading to the treasure. We employed easy communication trials in which communication went smooth, and challenging communication trials that contained staged misunderstandings. At the end of the hunt, the experimenter asked children a number of questions on their digital interaction partner, including questions from the areas of biology, communication, mental state and social relationship. These questions were used as a condition manipulation check (human vs. voice assistant condition), and crucially, to infer a parasocial score which was used to predict the adaptations in communicative behaviour. We measured children’s quantity of communication in terms of grounding utterances and number of syllables as well as children’s quality of communication in terms of pitch and intensity – in a baseline phase and three experimental phases (Figure 2).
In terms of quantity, children in the voice assistant condition decreased positive grounding and increased negative grounding in the challenging misunderstanding trials to a smaller amount than children in the human condition. In terms of quality, we found that while children who talked to a believed human digital partner, increased both f0 and intensity, in particular in challenging communication trials in which they tried to resolve the misunderstanding, no such adaptations occurred when children believed to interact with a voice assistant.
When children form stronger bonds with a digital interaction partner as measured by the parasocial score, they speak more, louder, and use more variable pitches. When a misunderstanding occurs, they are more engaged in resolving that misunderstanding and do not feel the need to adjust their communication strategy after one or two misunderstandings. When children form a weak social bond, they do not invest as much effort in repair strategies, but try to give more information after a misunderstanding and thus prevent another misunderstanding.
Our study provides a holistic approach to child-robot communication, taking both the quantity and quality of speech into account and investigating both easy and challenging communication situation. We discuss the relevance of our findings for developmental science at the interface to communication science and phonetics, as well as the direct practical implications that can be inferred for speech technology.

Authors

• Anja Gampe, University of Duisburg-Essen
  Presenting Author

• Katharina Zahner-Ritter

• Joanna Joys Müller

• Sarah Rebecca Schmid