| Abstract [eng] |
Speech emotion recognition problem which is considered relevant problem in human computer interaction area is tackled in this dissertation. Influence of human emotional condition to speech generation process is analysed, described feature selection process, reviewed feature selection methods, hierarchical classification is analysed and reviewed classification schemes, its experimental results in this dissertation. Based on research on feature selection methods the maximal efficiency criterion, minimal cross-correlation criterion, sequential forward selection technique are chosen and new hierarchical speech emotion classification scheme is proposed in this dissertation, which allows to use different feature sets in each classification stage. Experimental research on comparison of feature selection criteria shows that maximal efficiency, minimal cross-correlation, sequential forward selection provides similar results. No obvious advantage of either method can be stated after the research at the viewpoint of classification accuracy. Nevertheless, the Sequential forward selection technique enabled us to obtain up to 4 times smaller feature sets in comparison with other two techniques. The experimental research on newly proposed hierarchical classification scheme proves that the proposed scheme is more efficient than flat classification. For experimental purposes the classification problem of three (anger, joy, neutral), four (anger, joy, neutral, sadness) and five (anger, joy, neutral, sadness, fear) emotion classes were solved. The experimental results have proved hierarchical classification scheme advantage over flat classification – an improvement of 40 % in classification accuracy was achieved. |
