Audiovisual Affect Recognition in Spontaneous Filipino Laughter

Laughter has been determined as an important social signal that can predict emotional information of users. This paper presents an extension of a previous study that discovers underlying affect in Filipino laughter using audio features, a posed laughter database and categorical labels. For this study, analysis of visual (facial points) and audio (voice) information from a spontaneous laughter corpus with dimensional labels was explored. Laughter instances from a three test subject made up the corpus. Audio features extracted from the instances included prosodic features such as pitch, energy, intensity, formants (F1, F2 and F3), pitch contours, and thirteen Mel Frequency Cepstral Coefficients. Visual features included 170 facial distances taken from 68 facial points. Machine learning experiments were then performed in which Support Vector Machines -- Regression yielded the lowest mean absolute error rate of 0.0506 for the facial dataset. Other classifiers used were Linear Regression and Multilayer Perceptron.