Detecting laughter: Towards building a humorous machine

Abstract

Laughter is a common social gesture which often indicates the presence of humor. Detecting laughter and then understanding humor can make machines interact with us in a more natural way. This paper presents an algorithm to automatically detect laughter segments in speech. The voiced laughter of the speaker is recognized and the approximate onsets of the laughter bouts are used to annotate stored conversations. A simple algorithm based on the acoustic properties of voiced laughter is proposed and implemented for the same. The algorithm is able to detect the segments of laughter bouts in data consisting of sentences obtained from the switchboard corpus with an accuracy rate of 77.41% and a false detection rate of 12.90%.