Learning by Teaching: Designing Teachable Agents to Support Children's Pronunciation Skill Learning

Pronunciation is a critical yet challenging aspect of language acquisition, especially for children. In this work, we propose a new teachable agent system designed to support children's pronunciation learning through a Learning-by-Teaching paradigm to address this, featuring the Phoneme-level Mispronunciation Projection (PMP) method. This method enhances learning by having the agent reproduce or exaggerate children's mispronunciations, encouraging their correction and practice. We implemented a prototype system featuring both virtual and physical embodiments of teachable agents, allowing us to explore their effectiveness in motivating children. In a pilot study with 27 children aged 7-9, we evaluated the PMP method's perceived performance, children's motivation to teach, and the influence of agent embodiment (virtual vs. physical robots). Results demonstrate the effectiveness of PMP in supporting pronunciation learning and highlight the physical robot's advantages in fostering engagement. This work also offers initial design implications for future teachable agents to support children's pronunciation skill learning.