Fisher Feature Selection for Emotion Recognition

Emotion Recognition is one of the major areas of Affective Computing. Affective Computing is intended to reduce the communication gap between human and machine. Many research works in this area are focus on how to make a machine properly response to different mood of human. In this research, we propose Fisher Feature Selection (F-Score) for Emotion Recognition of Thai Speech to classify 4 different emotions from Thai utterance: Sad, Angry, Happy and Fear. The essence of our work lies on the inherit difficulties of Thai Language properties that intonation, articulation and accentuation can made different meanings. The proposed method is divided into two steps: First step, the human sound is extracted to get the 14 dominant features using F-Score Feature Selection. Then, in step two, two different learning networks are used to compare the classification performance. The results show that with the use of F-Score Feature Selection as a feature selection method to combine with Back Propagation Neural Network as a learning network offers a distinctive recognition rate of 95.13%.