LINGUISTIC COMPLEXITY MODERATES AI-MEDIATED SEGMENTAL FLUENCY GAINS: A QUASI-EXPERIMENTAL STUDY OF ELSA SPEAK IN PAKISTANI SECONDARY SCHOOLS

Among the most structurally under-developed areas of ESL teaching in Pakistani secondary schools, is pronunciation instruction, due to the large class size and lack of the teacher's phonological expertise, which makes it difficult for the learner to get individualized and real-time corrective feedback that is necessary for true phonological development in L2 pronunciation. This quasi-experimental study tested the effectiveness of AI pronunciation support using ELSA Speak in an instructional context based on TPACK in terms of the extent to which it led to more measurable segmental fluency improvements than traditional pronunciation instruction by a teacher — and whether the magnitude of the impact varied as a function of the level of the linguistic complexity of the task the students were asked to perform, ranging from word level to sentence level to para level production. Two hundred and forty ESL learners of secondary level (n = 60) from four schools in the private English-medium and government mixed-medium schools, both from Khushab and Jauharabad in Punjab Province, Pakistan were divided into equal groups for an experimental group (n = 60) who received twelve sessions of ELSA Speak 45 minutes each for four weeks and the control group (n = 60) who received regular classroom instruction on the same phonological content. The five PRAAT acoustic variables were assessed at three different levels, and segmental fluency was obtained. After normality testing by the Shapiro-Wilk test, Wilcoxon Signed-Rank and Mann-Whitney U tests were used. The experimental group showed outstanding within-group improvement, z = −4.78, p < .001, d = 4.97, or over seven times that of the average CAPT benchmark effect size of d = 0.79 (Lee et al., 2019). The between-group comparison strongly favored the experimental group (U = 17.00, z = −6.57, p < .001, d = 5.25). Analysis of these components individually showed that the magnitude of the gains was positively moderated by linguistic complexity; the effect size was increasingly higher from word level (d = 2.41) to sentence level (d = 2.43) to paragraph level (d = 4.42). The results expand the TPACK framework to incorporate AI-enabled pronunciation teaching, provide novel experimental backing for Skehan's (1998) automaticity model, and introduce new empirical data on the potential for equitable integration of mobile AI-assisted pronunciation instruction in both the school-based and non-school-based sectors in Pakistan. Keywords: The Effects of ELSA Speak on Segmental Fluency, Linguistic Complexity, AI-mediated Pronunciation and CAPT in Secondary School Students in Pakistan; Segmental fluency; Linguistic complexity; AI-mediated pronunciation; CAPT; TPACK; PRAAT; Quasi-experimental; Secondary education.