A speech singnal analysis method for automatic aggression detection in colloquial speech

Abstract

In recent years the destructive behaviour detection in the Internet task becomes more popular aiming to provide users’ psychological comfort. Destructive behaviour includes aggression which in European culture is presented as motivated destructive behaviour that can be oriented to the outside and the inside. Also, such behaviour contradicts currently accepted social norms. This paper implicates aggression as a paralinguistic phenomenon in the way that it reveals in speech and not what exactly was pronounced. This paper presents the definition and types of aggression as well as a short analysis of existing approaches for aggression detection in colloquial speech. The formalization of the multiclass classification task and description of the proposed approach also were presented in the paper. The experiments were made on the classification methods for automatic aggression detection, where the best result was achieved by the random forest. With the use of the random forest, we have got the best and most stable results. Based on the experiments the proposed approach for aggression detection was developed. Audio files from the mutlimodal corpora Stress at Service Desk Dataset and Aggression in Trains were used to train and test the models with the use of 5-fold cross-validation. The proposed approach includes an ensemble of random forests that were trained on different acoustic feature sets with different weights. The best result achieved using the proposed approach is 76.5 % in terms of unweighted average recall and is one of the best results achieved by other scientific groups.