Demonstrating Laughter Detection in Natural Discourses

This work focuses on the demonstration of previously achieved results in the automatic detection of laughter from natural discourses. In the previous work features of two different modalities, namely audio and video from unobtrusive sources, were used to build a system of recurrent neural networks called Echo State networks to model the dynamics of laughter. This model was then again utilized to detect laughters from presented test data. The approach was used to confirm human labels of laughter and to detect laughter in previously unmarked data, which resulted in nice results in offline applications. As reported in a publication currently under revision accuracies of 90 % were achieved using the multi modal input data relying on modulation spectral features from one microphone and movement data of a 360 degree camera positioned in the middle of a conference table [1]. In this work however, we would like to show a proof of concept for the online and on the fly recognition of laughter performing close to real-time. The goal of this work is to use a previously trained model of laughter in a modular process engine environment, which is currently under development overcoming known difficulties of pattern recognition and information fusion tasks, to detect laughter from a continuous microphone input. The intended application may then be used in online applications such as robots interacting with humans or the evaluation of human to human communication. Furthermore, the detection of laughter is an integral part of the improvement of dialog systems towards an affect understanding machine, since laughter is an important part of a healthy and natural communication.