ETRI extends its Korean AI tutor into reading and writing

South Korea’s ETRI said on 28 October that its prototype Korean-language tutor now lets learners discuss texts and receive automatic composition scores alongside speaking practice. A smart-glasses learning platform is also under development.[1]

The conversation system tracks lesson stages while showing learners only the topic. Familiar expressions go to specialised models with three billion to 13 billion parameters; unfamiliar ones go to a general model such as GPT. A further component filters repeated questions.[1]

This sits alongside a longer line of work on hearing and assessing learners’ speech. In a 2024 ETRI Journal paper, Kang and colleagues described the KSI Korean AI Tutor and EBS AI PengTalk. Their systems used speech recognition adapted to non-native speakers, then assessed pronunciation and fluency. The scoring model was trained to align its assessments with human experts’ judgements.[2]

That earlier paper separates two engineering problems clearly. A tutor must first recognise what a learner said, despite unfamiliar pronunciation or grammatical mistakes. It must then decide what feedback to give. Training the recogniser on learners’ speech addresses the first problem; learning a scoring rubric addresses the second.[2]

For the newer speaking system, ETRI cites an externally run, blinded assessment involving 102 people in late 2025. It reports appropriate-response rates of 93.8%, against 91.0% for GPT-4o-mini and 90.9% for GPT-5.1. The GPT systems performed better on expression variety and avoiding repeated questions.[1]

Those are conversation-quality findings, not measured language-learning gains. The current prototype’s broader lesson coverage is the new development.[1]

Report an error or suggest a correction