Determining Emotion Intensities from Audio Data Using a Convolutional Neural Network

Human beings communicate their feelings in the form of emotions. The feelings are expressed via speech, facial expressions, gestures, or other non-verbal signs. An emotion is a complex severe disturbance of an individual’s mental state that involves a subjective experience coupled with physiological, behavioral, and expressive responses. It denotes the mental state of the human mind and thought processes that represent a recognizable pattern. The emotions can be expressed with normal intensity or strong intensity depending on the incident being communicated. Detecting and classifying these emotions encompasses three fundamental machine learning processes; Feature Extraction, Feature Selection, and Feature Classification. Machine learning is the science of making computers learn and act like humans, while enhancing the learning with time in an independent manner, by giving them data and information in the form of observations and real-world interactions. A multi-modal approach comprising several machine learning algorithms is required to map out the intensities contained in the emotion classes. Mel Frequency Cepstral Coefficients are a set of about 10–20 features obtained from a speech signal describing the overall shape of a spectral envelope. The MFCC algorithm is designed to mimic the human hearing, thus it is ideal for this experiment.