Understanding the Student Confidence in Viva Responses Through Audio Analysis

Assessment of student responses during viva examinations presents challenges in terms of consistency, objectivity, and efficiency. Traditional methods rely on manual evaluation, which can be lengthy and prone to human bias. To address this, machine learning techniques are applied to automatically evaluate student confidence levels based on their oral responses. Student answers are transformed into numerical features using word embeddings such as TF-IDF, RoBERTa, BERT, GloVe, T5, GPT-2, FastText, and SBERT. A variety of machine learning classifiers, including MLP, AdaBoost, and Logistic Regression, were used to classify answers into confidence levels: weak, moderate, and strong. Among these, MLP with GloVe embedding models yielded the best results, achieving a test accuracy of 76%, an F1-score of 76%, and minimal overfitting with a training-test performance gap of 4%. These results demonstrate the strong potential of ML-based systems in improving automated grading and ensuring reliability and scalability in evaluating student responses.