Benchmarking classification models for emotion recognition in natural speech: A multi-corporal study

A significant amount of the research on automatic emotion recognition from speech focuses on acted speech that is produced by professional actors. This approach often leads to overoptimistic results as the recognition of emotion in real-life conditions is more challenging due the propensity of mixed and less intense emotions in natural speech. The paper presents an empirical study of the most widely used classifiers in the domain of emotion recognition from speech, across multiple non-acted emotional speech corpora. The results indicate that Support Vector Machines have the best performance and that they along with Multi-Layer Perceptron networks and k-nearest neighbour classifiers perform significantly better (using the appropriate statistical tests) than decision trees, Naïve Bayes classifiers and Radial Basis Function networks.