Automatic Analysis of Emotions from the Voices/Speech in Spanish TV Debates

The goal of this work is to automatically analyze the emotional status of speakers, in human-human interactions, carried out in TV debates, where controversial topics are often presented. Human observers provide their perception about the emotional status associated to the interventions of the participants. An analysis of the resulting annotation was carried out by using different models for representing the emotions. The obtained labeled corpus was used to build an automatic system capable of detecting the emotional status associated to each acoustic signal, making use of the deep learning paradigm. The use of a corpus, where the real emotions that appear in a Spanish TV debate (with subtleties and often closer to neutrality than acted ones), are represented is crucial for learning models properly. In fact, although the level of accuracy depends on the problem complexity and the model employed for representing the emotional status, F1 scores of 0.7 were attained.