Acoustic and Signal-Based Evaluation of AI-Assisted English Speech Training for Arabic-Speaking University Learners

The biomechanical process that produces human speech creates measurable acoustic signals by vocal folds vibration and vocal tract resonant filtering. This study analyzed sound and vibration principles to estimate the effectiveness of Artificial Intelligence (AI)-assisted pronunciation training in a university context. The study recorded speeches for 200 Arabic-speaking undergraduate students studying a general English course. Students were assigned randomly to an experimental group using ELSA Speak (a mobile AI application) that provides pronunciation feedback based on acoustic features and a control group receiving traditional teaching methods without using AI support. Researchers used a pre-test/post-test experimental design and used Praat and MATLAB to analyze speech recordings from reading and speaking tasks to extract objective acoustic measures, including fundamental frequency (F0) stability (jitter), formant frequency consistency (F1, F2 dispersion), and spectral characteristics (spectral centroid and bandwidth). The results of the paired t-tests and ANOVA showed that the experimental group achieved significant improvements (p < 0.05) compared to the control group. Specifically, pitch variability decreased noticeably (F0 standard deviation reduced by 39.5%), vowel articulation became more stable (formant dispersion reduced by 30–40%), and spectral organization improved. The findings also found that AI-assisted pronunciation training leads to measurable improvements in speech signal stability and clarity. This study highlights the value of sound analysis as a quantitative framework for evaluating human-centered speech learning systems.