Identificação de sentimento em voz por meio da combinação de classificações intermediárias dos sinais em excitação, valência e quadrante

Abstract

Speech emotion recogntion is important in areas such as health, psychology, and telemedicine for information about an individual’s states of emotions. Speech emotion recogntion is commonly performed in categorical classes, such as “sadness” or “joy”. According to Rus sell’s map of affection, emotions can also be classified by arousal (excitation), valence, and quadrants. In this work is proposed a method to increase the performance of speech emotion recogntion in categorical classes using classifiers that perform intermediate classification in the classes of valence, excitation and quadrants using a multiview approach. Moreover, three types of classifiers perform the same task, using different features extracted from the voice signal, which combine in one Ensemble they tend to increase individual results. To combine these results and obtain the final classification, a decision tree is proposed and that increases F1 metrics from 0.61 by Ensemble of three kinds of classifiers to 0.63 in a public database