Measuring Spoken English Proficiency Level Based on IELTS Speaking Test Using Machine Learning Models

This study presents the development of a system for evaluating English-speaking proficiency based on the IELTS framework. We employed data collection, analysis, and machine learning techniques to build a predictive model. The system focuses on using the statistical features extracted from the transcription of the speaking test. The initial model (XGBoost) showed a mean squared error of 2.29. After feature reduction and hyperparameter tuning, performance improved, achieving an average cross-validation score of 1.57 with a variance of 0.3872 across five folds. Our analysis found that lexical diversity, measured by the total number of unique words, is the most influential factor in predicting IELTS scores, while other features had minimal impact. Despite the improvements, the model does not account for critical speech components like fluency and pronunciation. Future research should address these aspects to better align with IELTS criteria. This study contributes to natural language processing and linguistics by offering a new machine learning model for predicting IELTS scores from the test transcription.