Degrees of Anger Prediction from Speech

In recent times, a huge body of research has been conducted for speech processing applications, particularly speech recognition. However, a few studies have been found to predict users’ anger (i.e., yes/no) from their speech. In this study, we predict users’ degrees of anger (i.e., minimal, mild, moderate, and extreme) based on their speech which has not been studied yet. We apply different machine learning-based approaches for obtaining better accuracies. We extract important features from speech and apply a feature selection technique. We apply machine learning models like Support Vector Machine (SVM), Gaussian Naive BayeS (GNB), and Multinomial Logistic Regression (MLR) along with ensemble models like Random Forest (RF), Extreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), and Bagging to predict the degrees of anger from speech. Among the models we have used, XGBoost demonstrates an F1 score of 93.2% which outperforms the state-of-the-art models. To the best of our knowledge, this is the first attempt to predict the distinct degrees of anger of a person by using his speech sample unlike the frequently used binary classification o f a nger states.