Poster Presentation AI-Generated Elicited Imitation Materials for EFL Speaking: Comparing three LLMs more

Japanese EFL learners struggle with speaking comprehensibility and intelligibility, and while High Variability Phonetic Training (HVPT) and Elicited Imitation (EI) show promise, creating level-appropriate EI materials at scale requires efficient methods. This study investigates whether AI can generate appropriate EI sentences for CEFR A2-B1 learners by comparing three AI models (ChatGPT-5, Claude-Sonnet-4, Gemini-1.5-Flash) generating EI sentences through systematic prompt engineering. Materials (n=150 per model, 50 per CEFR level) were evaluated by three expert raters on grammatical complexity, vocabulary appropriateness, and sentence length suitability. MANOVA revealed significant multivariate effects (Wilks' λ = 0.731, F(6, 888) = 23.18, p < .001, η²p = .145), with Claude significantly outperforming other models on vocabulary appropriateness (F(2, 447) = 42.67, p < .001, η²p = .160) and sentence length suitability (F(2, 447) = 35.24, p < .001, η²p = .136), while showing no differences in grammatical complexity. Results establish Claude as optimal for generating level-appropriate speaking materials, enabling scalable, personalized EFL assessment tools. This advances ML in CALL by demonstrating systematic AI model comparison for pedagogical material generation, supporting subsequent phases investigating HVPT's impact on speaking performance.