Artificial Intelligence in English Language Assessment Transforming Traditional Oral Exams into Adaptive Learning Tool

The evaluation of spoken language proficiency plays a crucial role in language learning and assessment systems. Existing methods, such as those relying solely on acoustic or textual features, often fail to fully capture the nuances of speech, leading to suboptimal performance. The present research extends the existing approach in identifying impaired language by presenting a new model that incorporates the text and acoustic analyses simultaneously. RoBERTa, a transformer-based architecture is used to characterize speech transcripts, and Melfrequency cepstral coefficients (MFCCs) incorporate essential acoustic features of the signal including tone and pitch. These features are then fused to give a single representation which is then classified using an optimized XGBoost model. It operates to optimize the model’s performance, the goal of hyperparameter tuning that it can reach precise classification accuracy. The experimental outcomes confirm RoBERTa-XGBoost model by enhancing the accuracy to 92.5%, precision 91.1%, recall 93.4%, and F1 score of $92.3 \%$ as opposed to other existing models. This model can be used to produce an improved system for the assessment of language proficiency as it considers both textual and acoustic attributes.