The decisive emotion identifier?

Abstract

Emotion recognition from speech is a relatively new research area with wide applications such as patient monitoring, call centers and human-robot interaction etc. A number of methods such as SVMs, GMMs, HMMs etc have been used in the past for emotion recognition. This paper describes an experimental study on four basic human emotions namely anger, happiness, sadness and neutral. An emotional database is formed by the recording one word utterance `Hello'. Pitch, energy and TILT parameters are the basic features used for the detection of emotion. The Classification and Regression Tree (CART) called wagon is used as a classifier to train and test the type of utterances within the four categories. This paper tests the ability of the classifier to categorize the emotion on the basics of individual as well as combination of different features with each other using the CART classifier. The emotional recognition accuracy of these experiments allows us to compare the emotional information contained by each feature. Finally, we suggest the best combination of features which gives the highest accuracy for recognition.