State of the Art Recognition of Emotions from Speech, Using a Low Complexity Solution Based on Reaction-Diffusion Transform

A low complexity solution for the problem of emotional speech recognition is provided and evaluated demonstrating state of the art accuracies for relevant emotional speech databases. The relevance of proper augmentation in the sound domain is demonstrated, with important increases in accuracies. In comparison to more traditional methods based on mel-cepstral coefficients or recurrent layers, our method offers a simpler way to compute spectrograms and consequently would reduce the complexity of the implementation, for instance the proposed method was recently implemented in a mobile platform (running autonomously from any network connectivity). Test accuracy of 96.83 % is obtained for the EMO-Db dataset and 96.63% for the RAVDESS datasets.