Publication

Abstract : Speech emotion recognition system is a discipline which helps machines to hear our emotions from end-to-end. It automatically recognizes the human emotions and perceptual states from speech. This work presents a detailed study and analysis of different machine learning algorithms on a speech emotion recognition system (SER). In prior studies, single database was experimented with the sequential classifiers to obtain good accuracy. But studies have proved that the strength of SER system can be further improved by integrating different deep learning classifiers and by combining the databases. Model generalization is difficult with a language-dependent and a speaker-dependent database. In this study, in order to generalize the model and enhance the robustness of SER system, three databases namely Berlin, SAVEE, and TESS were combined and used. Different machine learning paradigms like SVM, decision tree, random forest, and deep learning models like RNN/LSTM, BLSTM (bi-directional LSTM), and CNN/LSTM have been used to demonstrate the classification. The experimentation result shows that the integration of CNN and LSTM gives more accuracy (94%), when compared to other classifiers. The model performs well in all emotional speech databases used.