Segment based emotion recognition using combined reduced features

Abstract

The attitude of a human being involves with their emotions. Emotions can be observed in either verbally or visually or both. Verbal emotion recognition is a difficult task and an area of speech processing. It has a wide variety of applications in almost all fields. In this work, the authors have tried to recognize five types of emotion as anger, sadness, happiness, fear, and neutral. The work is focussed on the choice of spectral feature computation. For such purpose, Mel-frequency Cepstral coefficients (MFCC), spectral roll-off, spectral centroid and spectral flux are considered on frame-level extraction. Some of these features need to be reduced, combined, and balanced. The combined methods are verified and observed the effectiveness of results. The resulting features are used with neural network (NN) based models for recognition purpose. The models of multilayer perceptron (MLP), radial basis function network (RBFN), probabilistic neural network (PNN) and deep neural network (DNN) are considered and tested for the chosen features. It is observed that less amount of features provides reliable accuracy in case of PNN. The same utilizes less time for training and testing in case of MLP, RBFN, and PNN. However, DNN is not suitable for fewer amounts of features. It requires large data for better accuracy in the particular field. The results support the PNN with an average accuracy of 96.9% with low-dimensional feature sets, whereas the average accuracy of MLP, RBFN, DNN models found 90.1%, 92.7%, and 73.6% respectively.