Speech-based emotion recognition of the distant student with adaptive intellectual information technologies

To overcome the negative sides of distance education and to develop the communicative competence of the distant student, it is necessary to improve the interactive dialogue systems, in particular, to realize the opportunity of speaker state recognition. Although lots of excellent results have already been achieved in this sphere, there are some open questions. Recently scientists have developed various program systems which are good at extracting numerical characteristics from speechsignals. Unfortunately, the amount of features might be huge that becomes a challenge for classification models. Therefore it is significant to determine relevant features from data sets. In this paper we consider the feature selection procedure that is based on the adaptive multi-objective evolutionary algorithm and investigate its efficiency in combination with different classification models. Generally, the feature selection procedure can be organized as a wrapper approach or a filter one. Compared with the wrapper approach, the second technique requires less computational resources and also demonstrates good results. Therefore in this research we implemented the feature selection procedure according to the scheme of the filter approach. Furthermore, to avoid choosing the genetic algorithm settings we developed a self-adaptive modification of the conventional multi-objective genetic algorithm. Due to application of the self-adaptive heuristic optimization procedure it became possible not only to improve the performance of involved classifiers but also to reduce the number of selected features essentially. Obtained results demonstrate high performance of the developed algorithmic scheme and imply the reasonableness of its usage in the dialogue system modules for recognition of student emotions distantly.