Differential responsiveness to AI-mediated feedback in EFL segmental pronunciation learning

Abstract

This study examines changes in recurrent English segmental errors during an AI-mediated pronunciation feedback condition. Although recurrent segmental difficulties among Vietnamese EFL learners are well documented, less is known about how these errors change during AI-mediated practice or how learners work with the feedback they receive. The study used a quasi-experimental mixed-methods design with 105 first-year English majors at a Vietnamese university. The experimental group (n = 53) completed three weeks of AI-assisted pronunciation training with ELSA Speak as the feedback environment; the control group (n = 52) continued with conventional classroom-based practice. Overall segmental error rates fell by 13.3 percentage points in the experimental group and by 4.1 points in the control group. The post-test difference between groups was statistically significant (p < .01), with a large recorded effect size (Cohen's d = 0.85). Change also varied by error category. Final consonant omission showed the clearest descriptive reduction, whereas change in the /ʌ/-/æ/ and /e/-/eɪ/ distinctions was less even. In the interviews, learners described several ways in which the automated feedback shaped their practice. They reported noticing errors that had previously escaped their attention, trying the same item repeatedly, and practicing outside scheduled class time. Yet the feedback was not always easy to interpret. Several participants found it difficult to determine what the system was identifying when the pronunciation problem was more complex. These results suggest that AI-mediated pronunciation feedback is best understood selectively rather than as uniformly effective across targets or learners. In this study, the condition was associated with larger overall gains, but category-level change remained uneven, and the practical value of the feedback depended partly on learners' ability to understand and act on it.