Feature Analysis for Speech Emotion Classification

The automated classification of speech emotions is a potential candidate for clinical applications or even for educational purposes in the training of students. A description of procedures to evaluate emotional states from voice recordings in different environments is presented, together with extraction and the selection of features such as Fundamental Frequency, energy, formants and Mel-Frequency Cepstrum Coefficients (MFCC). For comparison purposes three speech corpus, one made by actors using the German language, and two made by inducing emotions in English and Brazilian Portuguese. A number of 208 features were extracted, this number was reduced using selection and emotion classification was performed by the use of a Support Vector Machine algorithm. The objective of this work was to compare results from different databases based on the classification of emotions with supervised learning algorithms.