Smart computers : can you tell my mood ?

Abstract

This project aims to build and develop a classifier that is able to sense the emotion of a person based on audio signals. The ability for computers to understand a person’s emotions can be found in applications such as a lie detector, and call centers. The classifier is built and designed on MATLAB, based on multi-layer perceptron with back propagation as the training function. Speech features that can be used for emotion classification will be studied. The database that is used for this classification project is the Berlin Database of Emotional Speech. A systematic process is taken to optimize the accuracy of the classifier. Different parameters will be changed with a set of input features to obtain the optimal accuracy on the dataset. For different combinations of parameters, analysis will be provided to justify the performance. Subsequently, the classification model is then packaged into a Graphical User Interface (GUI) as the final product. Experimental results showed that only certain features contribute more towards the accuracy of the classifier. Also, emotion classes that have similar intrinsic characteristics are often confused with each other.