RNN-Based Method for Classifying Natural Human Emotional States from Speech

Abstract

An accuracy of classifying human emotional states from speech has been dependent on emotional speech database, speech signal processing algorithms, and classification methods. Numerous static and dynamic classification techniques are used for emotional state recognition. The paper considers a novel method for classifying natural positive, negative, and neutral human emotional states based on a five-layer Recurrent Neural Network (RNN). The method novelty is due to various neuron activation functions employed for each network layer stemming from the peculiarities of natural emotional speech informative parameters used as input data for the neural network. Local and global informative speech parameters relevant to human emotional states are outlined, and the known classification methods are surveyed. The proposed classification method is described, and the developed Russian database of natural emotional speech is presented. The research results compared with the widely used classification methods have evidenced 95 % accuracy of the developed method, which can be effectively tested in systems for detecting and classifying natural human emotional states from speech.