Authors - Mahimaben Panjabi, Vijay Makwana Abstract - This study investigated the efficacy of Large Language Models (LLMs) for providing formative assessment of spoken English within a mixed-methods, one-group pre-test/post-test design. The research involved 30 intermediate English as a Second Language (ESL) learners and five expert faculty members. The design integrated quantitative pre-test/post-test data with qualitative analysis of AI-generated feedback, student reflections, and faculty interviews. During the intervention, all students used LLM tools (ChatGPT and Google Gemini) for practice and feedback. To ensure a robust, triangulated assessment, a coder-based framework was implemented where human experts evaluated spoken tasks against linguistic criteria. A paired-samples t-test revealed a statistically significant improvement in speaking proficiency from pre-test to post-test. Thematic analysis of qualitative data indicated that LLM-generated feedback was perceived as useful, accessible, and effective in reducing learner anxiety. This multi-faceted approach affirms that LLMs can be effective supplementary tools in language education and offers a holistic model for their integration into formative assessment practices.
