Characterization and Classification of Speech Emotion with Spectrograms

Abstract

The work attempts to characterize and classify speech emotions using the spectrogram. Initially, it extracts the individual Red, Green, and Blue parameters from the raw speech spectrogram image of every individual emotional utterance. Further, it computes the statistical parameters of individual RGB components to characterize the chosen emotional states. The utterances of anger, happiness, neutral, and sad emotional states from the standard Berlin (EMO-DB) database has been used for this purpose. The individual statistical R, G, and B spectrogram parameters are found to be different within an emotion as well as across emotional states. Thus, these values have been used as different feature sets to classify the designated emotional states using the popular Multilayer Perceptron Neural Network (MLPNN).