Classifier Fusion Based on Gaussian Mixture Models (GMM) and Support Vector Machine (SVM) in the Function of Classification Emotional Speech, Using Spectral and Prosodic Features

Various types of information are imprinted in the speech signal, which the interlocutor registers with his auditory perceptual system. The number of transmitted messages, hidden in the speech signal, is numerous, because they give a stamp on the complete personality of the speaker: his biological, sociological, psychological and social dimensions. Studying and analyzing the emotional characteristics of the speech signal requires knowledge of several scientific fields, such as information technology, psychology, pedagogy, acoustics, signal processing, linguistics, medicine, artificial intelligence and many others. In the process of processing human speech, in order to determine the emotional state of the speaker, in this paper we used its spectral and prosodic characteristics. As the emotional state of the speaker causes a change in the vocal tract in relation to neutral speech, a change in rhythm, intonation, melodicity, volume and other changes, i.e. adequate changes in the spectral plane, we believe that these characteristics are relevant in the speech signal for determining the emotional state of the speaker. Also, prosodic characteristics related to rhythm, accent, intensity and intonation of speech [3] largely reflect emotional speech. In order to classify the speech signal into a specific emotional group, a fusion of classifiers based on Gaussian Mixture Model was used (GMM) and Support vector machine (SVM). SAVEE emotional speech database was used for training and testing data.