Language-agnostic speech anger identification

Following the constantly increasing adoption of affective computing based solutions, this paper investigates the feasibility of multilingual anger identification. To this end, we formed such a corpus by suitably combining seven different datasets representing five different languages, i.e. English, German, Italian, Urdu, and Persian. After analyzing the diverse characteristics of the datasets, we designed four classification algorithms, namely Support Vector Machine, Decision Tree-based Bagging scheme, Convolutional Neural Network, and Convolutional Recurrent Neural Network. Such classification mechanisms are trained on appropriate features extracted from time and/or frequency domains, while speech data have been balanced considering every diverse characteristic incorporated in the datasets (language, sex, acted, etc.). Our findings render multilingual anger identification feasible since the proposed audio pattern recognition methodology based on Mel-spectrograms and CRNN achieved quite satisfactory identification rates.