Predictive Analysis of Confidence and Clarity based on Features of Audio and Transcribed Text

Viva examinations play a crucial role in assessing a candidate’s confidence, language proficiency, and communication skills. However, traditional viva assessments often lead to subjective, qualitative results that can vary between examiners. The paper presents a machine learning-based approach to objectively evaluate the confidence and clarity of viva answers, aiming to establish a more consistent and reliable assessment system . The proposed system uses NLP to quantify aspects like tone, pace, and choice of words used by the respondent to arrive at confidence levels. Sense-making is also used to assess the substantiation of the answers besides checking the grammatical correctness and the logical flow of the answers. The best results were achieved using stacking model(combining MLP, KNN and Random Forest with a AdaBoost as the meta-model) for confidence classification with an F1-score of 92.8% and a stacking model (combining MLP and Random Forest with a Decision Tree as the meta-model) for clarity classification with an F1-score of 91.4%