An Approach towards Emotion Recognition from Speech using Deep Learning

The significance of emotions in human interactions and decision-making necessitates the development of advanced techniques for emotion recognition, particularly from speech signals. In this research, deep learning models and paralinguistic features are implemented for the purpose of classifying emotions which are communicated through speech. Initially, a specific set of features is extracted from speech signals and subsequently, a deep convolutional network with auto-encoders is employed to transform these features into a higher-level representation. For the classification task, a multilayer perceptron (MLP) classifier trained to identify six fundamental emotions (happiness, surprise, sadness, fear, anger, disgust) is utilized. The proposed methodology is assessed through a five-fold cross-validation strategy using the eNTERFACE'05 dataset, achieving an accuracy of 91.00% and surpassing the performance of various state-of-the-art methods.