A Hybrid Distance-Based Method and Support Vector Machines for Emotional Speech Detection

We describe a novel methodology that is applicable in the detection of emotions from speech signals. The methodology is useful if we can safely ignore sequence information since it constructs static feature vectors to represent a sequence of values; this is the case of the current application. In the initial feature extraction part, the speech signals are cut into 3 speech segments according to relative time interval process. The speech segments are processed and described using 988 acoustic features. Our proposed methodology consists of two steps. The first step constructs emotion models using principal component analysis and it computes distances of the observations to each emotion models. The distance values from the previous step are used to train a support vector machine classifier that can identify the affective content of a speech signal. We note that our method is not only applicable for speech signal, it can also be used to analyse other data of similar nature. The proposed method is tested using four emotional databases. Results showed competitive performance yielding an average accuracy of at least 80% on three databases for the detection of basic types of emotion.