AI-Mediated Speaking Practice for Aptis ESOL Preparation: An Explanatory Sequential Mixed-Methods Case Study of Vietnamese EFL University Students

Artificial intelligence is increasingly used in English language learning, yet empirical evidence remains uneven regarding how AI-mediated speaking practice supports performance in standardised speaking-test preparation. This explanatory sequential mixed-methods case study examined how Vietnamese EFL university students preparing for Aptis ESOL Speaking engaged with an AI-supported speaking programme. Forty-eight students participated in an eight-week intervention integrating Aptis-style speaking tasks, ChatGPT voice interaction, automatic speech recognition (ASR) feedback, teacher scaffolding, and reflective journals. Quantitative data included pre/post Aptis-style speaking scores, speaking anxiety ratings, willingness-to-communicate scores, and AI-use logs. Qualitative data were collected from semi-structured interviews, learner journals, and teacher field notes. The reported findings indicated a statistically significant increase in speaking performance from pre-test to post-test, alongside reduced speaking anxiety and increased willingness to communicate. Qualitative findings suggested that AI created a low-stakes rehearsal space, increased opportunities for output, supported lexical and pronunciation noticing, and enhanced learner autonomy. However, students also reported over-reliance on AI-generated scripts, uneven feedback quality, and difficulties transferring rehearsed fluency to spontaneous test performance. The study argues that AI is most pedagogically valuable when integrated as a scaffolded rehearsal partner rather than as a substitute teacher, examiner, or answer generator.