EXPLORING PATHWAYS OF UNIVERSITY STUDENTS’ PERSONALISED ENGLISH-SPEAKING LEARNING WITH GENERATIVE ARTIFICIAL INTELLIGENCE

Abstract

English-speaking instruction has long been plagued by insufficient input, lack of interaction, delayed feedback, and emotional anxiety, leading to unsatisfactory outcomes. Generative Artificial Intelligence (GenAI) offers a new solution to these challenges by enabling real-time, adaptive conversational practice. This study systematically examined the potential effectiveness of GenAI in enhancing learners' English-speaking proficiency using a mixed-methods approach that combined pre- and post-test speaking assessments, weekly reflective journals, and semi-structured interviews. Three customised agents: an AI Reading Assistant, an AI Speaking Partner, and an AI Speaking Examiner; were integrated into participants' learning routines. The findings revealed that GenAI contributed to improvements in pronunciation, grammatical accuracy, speaking fluency, and learners’ confidence by offering immediate feedback and a low-anxiety learning environment. Participants also reported increased willingness to communicate and greater opportunities for repeated speaking practice in a supportive environment. In addition, GenAI-assisted interaction encouraged learners to engage in more self-directed speaking activities and facilitated greater awareness of their language problems during oral production. However, variations were observed across learners with different proficiency levels, particularly in the development of specific speaking micro-skills. Despite its pedagogical potential, GenAI also demonstrated limitations, including weak memory continuity, limited emotional authenticity, and difficulties handling culturally nuanced or complex communicative contexts. Overreliance on AI-generated responses during speaking practice may reduce opportunities for spontaneous language production and for independent use of communicative strategies. In addition, the accuracy and relevance of AI-generated feedback occasionally varied, leading to uncertainty among learners when evaluating and revising their spoken output. The results offer new insights into English-speaking instruction, emphasising the need for adaptive prompt engineering and teacher–AI collaboration. They also provide guidance on improving the use of GenAI in speaking instruction, such as designing scaffolded tasks for diverse proficiency levels.