“You said da da...”: A Short Echoing Tweak for Journaling with VA
Jisu Ryou, Kihun Lee, Joongseek Lee · 2023
AI and deep learning have enabled development of voice agents (VAs) and conversational agents (CAs) for more varied uses, including audio journaling which requires longer, more nuanced conversations. Current technology is capable of utilizing additional data, but VAs still remain limited to single-turn, task-oriented interactions. In this study, we investigate whether current turn-taking system is suitable for interactions with long speech in the form of audio journaling, and suggest ways that would make VAs more appropriate to handle longer sentences and turns by the user. To this end, we designed three types of responses, Empathetic, Reading-back, and Mixed, which all echo the user’s speech in different ways. From a week-long journaling experiment with 23 subjects, we collected long speech data (average number of syllables: Empathetic=425, Reading-back=438, Mixed=304). The most effective response in this case is a Reading-back in terms of quality of data and journaling experience.