Evaluating automatic laughter segmentation in meetings using acoustic and acoustic-phonetic features

In this study, we investigated automatic laughter segmentation in meetings. We first performed laughterspeech discrimination experiments with traditional spectral features and subsequently used acousticphonetic features. In segmentation, we used Gaussian Mixture Models that were trained with spectral features. For the evaluation of the laughter segmentation we used time-weighted Detection Error Tradeoff curves. The results show that the acousticphonetic features perform relatively well given their sparseness. For segmentation, we believe that incorporating phonetic knowledge could lead to improvement. We will discuss possibilities for improvement of our automatic laughter detector.