Automatic video segmentation tool for laughter detection based on audio features

Non-linguistic signals, specifically, laughter offers a lot of information such as cues on the emotional state of a person a topic changes in meetings. The numerous benefits of laughter, ranging from identifying activities to improving speech-to-text accuracy, have gained the interests of many researchers. Most of the recent studies regarding laughter have successfully and effectively detected laughter using different modalities. However, they have encountered several problems in manually segmenting laughter as there is no automated segmentation tool for video containing laughter. This paper presents a prototype that would automatically segment laughter from videos of meetings. The prototype uses a segmentation algorithm pattered over the model-based segmentation approach. A SVM trained on audio features (such as MFCC, Pitch, and Formants) to classify instances. In addition, this research involves the creation of a Filipino Meeting corpus which consists of videos of spontaneous meetings. The classifier achieved an accuracy of over 90%. The segmentation algorithm achieved AUC-ROC of 0.79. The prototype is able to accurately segment laughter in videos however, errors are encountered (which is usually composed of onset and offsets of laughter and other sound accompanied with breathing). In addition, the research was also able to characterize the laughter models built and define laughter in the corpus