Emotion recognition from telephone speech using acoustic and nonlinear features

This paper addresses the problem of the automatic recognition of emotional states from speech recordings, especially those kind of emotions reflecting that the life or the human integrity are at risk. The paper compares the performance of two different systems: one being fed with speech signals recorded directly from the people (whole spectrum) and other one in which the speech signals are recorded through a telephone channel. The characterization stage is based on cepstral, noise and nonlinear features, and the classification strategy uses a fusion of multiple classifiers (Gaussian Mixture Models - Universal Background Model and Support Vector Machines). The proposed system achieves classification rates around 99%, even in the case of telephone speech.