Development of Natural Emotional Speech Database for Training Automatic Recognition Systems of Stressful Emotions in Human-Robot Interaction

摘要

A comparative analysis of 18 popular speech databases used in the field of affective computing was carried out. A database containing unique audio recordings of Russian speech (words, phrases, and sentences), registered in natural experience of emotions by eyewitnesses and/or participants in adverse events under stress, has been developed. Voice recordings were taken from video materials of disasters and emergencies occurred on the territory of Russia in 2015-2019. Examination results of psychological and psychiatric reactions of 293 eyewitnesses and/or participants (119 men and 174 women, aged 15 to 65 years) to traumatic factors of adverse events are presented. In accordance with the development results, the following key conclusions have been made: in the future, the database will allow researchers to objectively assess the capabilities of existing methods and approaches for analyzing emotional speech, as well as to assess the need for new mathematical devices in speech signal processing; the voice database can be successfully used as training or test sample, necessary for effective training of systems for automatic recognition of stressful emotions in human-robot interaction.