Deep Learning with Virtual Agents: How Accented and Synthetic Voices Affect Outcomes

The current study investigates accent effects using virtual agents in the context of a multimedia learning environment. In a 2 (voice type: human, synthetic) x 2 (voice accent: English, Russian) between-subjects factorial design, the source and accent of the agent’s voice were manipulated. Research has shown that an instructor’s accent can have an impact on learning outcomes and perceptions of the instructor. However, these outcomes and perceptions have yet to be fully understood in the context of a virtual human instructor. Outcome measures collected included: knowledge retention, knowledge transfer, and cognitive load. Perception measures were collected using the Agent Persona Instrument-Revised, API-R, and a speaker-rating survey. Overall, there were no significant differences between the accented conditions. However, the synthetic condition had significantly lower knowledge retention, knowledge transfer, and mental effort efficiency than the professional voices in the human condition. Participants rated the human recordings higher on speaker-rating and API-R measures. These findings demonstrate the importance of considering the quality of the voice when designing multimedia learning environments.