Speech-Based Emotion Classification for Human by Introducing Upgraded Long Short-Term Memory (ULSTM)

Abstract

All humans have the intelligence for emotions through emotional behaviour by social skills and by interacting with and imitating humans. Not only that, we also enhance and upgrade our skills for analysis of different emotions through learning with our experience in our surroundings. Now, what if the machine is capable of learning through its artificial intelligent skills? The ongoing exploration is being done using the deep learning model concept. This technique is being used to enhance the learning capacity of the machine which is most important in human emotion recognition because one emotion can be derived towards another type of emotion which is difficult to analyse. This theme has inclined us to explore this problem. The proposed method has been designed to categorize the human emotions through four different deep learning models, which are convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM) and gated recurrent unit (GRU). For training these models, well-known physical and perceptual features have been fitted. The system has been tested on the benchmark data of Ryerson Audio-Visual Dataset for Emotional Speech and Song (RAVDESS). Furthermore, the mentioned deep learning model has been compared based on testing the above dataset in terms of the vanishing gradient problem. In addition, an upgraded model of LSTM has been proposed to get better accuracy and it is being tested with the existing model of LSTM.