Developing a Framework for Metrics and Evaluation of the Impact of Acoustic-Prosodic Features in Synthesized Speech on Listener Perception in Dyadic Interactions

Abstract

Acoustic-prosodic properties of conversational speech have been shown in prior research to impact the perceptions that listeners have towards the speakers in dyadic interactions. While correlations between the two have been consistently found, no standardized framework yet exists for systematically assessing the nature of such relationships in an easily reproducible fashion. The purpose of this thesis was to develop an experimental software prototype that facilitates researchers in their collection and assessment of data to explore the relationships between the audio qualities of a conversational dialogue agent’s synthesized responses entrained on the acoustic-prosodic features of their human interlocutor’s speech, and the way human listeners perceive that dialogue agent. Experiments were run to replicate results from past studies using the software developed.