Robust and reliable techniques for speech-based emotion recognition

One of the crucial challenges related to the spacecraft control is the monitoring of the mental state of crew members as well as operators of the flight control centre. In most cases, visual information is not sufficient, because spacemen are trained to cope with feelings and not to express emotions explicitly. In order to identify the genuine mental state of a crew member, it is reasonable to engage the acoustic characteristics obtained from speech signals presenting voice commands during the spacecraft control and interpersonal communication. Human emotion recognition implies flexible algorithmic techniques satisfying the requirements of reliability and fast operation in real time. In this paper we consider the heuristic feature selection procedure based on the self-adaptive multi-objective genetic algorithm that allows the number of acoustic characteristics involved in the recognition process to be reduced. The effectiveness of this approach and its robustness property are revealed in experiments with various classification models. The usage of this procedure leads to a reduction of the feature space dimension by a factor of two (from 384 to approximately 180 attributes), which means decreasing the time resources spent by the recognition algorithm. Moreover, it is proposed to implement some algorithmic schemes based on collective decision making by the set of classifiers (Multilayer Perceptron, Support Vector Machine, Linear Logistic Regression) that permits the improvement of the recognition quality (by up to 10% relative improvement). The developed algorithmic schemes provide a guaranteed level of effectiveness and might be used as a reliable alternative to the random choice of a classification model. Due to the robustness property the heuristic feature selection procedure is successfully applied on the data pre-processing stage, and then the approaches realizing the collective decision making schemes are used.