Research into the Influence of Informative Speech Parameters on the Reliability of RNN-based Classification of Human Emotional States

Reliable classification of human emotional states from speech is dependent on speech signal processing algorithms and classification methods. The reliability is especially affected by the database of emotional speech informative parameters used to train classification methods. The paper is aimed at studying the influence of informative speech signal parameters on classification reliability of human emotional states based on a Recurrent Neural Network (RNN). The research is to expand the information space for emotional speech parameters through the use of an original signal representation algorithm based on Empirical Mode Decomposition (EMD), and assessing the impact on classification reliability using a five-layer RNN. The article outlines the research relevance and provides a brief overview of related academic papers. Local and global informative speech parameters relevant to human emotional states are considered in detail. The emotional speech database is described, and the speech signal representation algorithm based on modified decomposition methods is characterized. The research results on the method for classifying human emotional states based on a five-layer RNN using original and extended databases of emotional speech informative parameters are presented. Reliability assessment was performed by computing the confusion matrix for multi-class classification. The research results have evidenced an increase in classification reliability of human emotional states by an average of 2.66 % through expanding the information space for emotional speech parameters. The reliability ranges from 97.4 % to 98.3 % for the classification of emotional states, and from 95.8 % to 98.2 % for the classification of emotions.