Acoustic differences between laughter and screams in spontaneous dialog

抄録

This study investigates acoustic differences between laughter and screaming in spontaneous dialog. First, three classification experiments, i.e., speaker-closed, speaker-open, and corpus-open experiments, demonstrate the acoustic discriminability of spontaneous laughter and screaming. The results of the speaker-closed experiment reveal that laughter and screams can be acoustically classified with high F-measures of 91.80% and 82.39%, respectively. The results of the speaker-open experiment reveal that the model exhibites 2.27% (laughter) and 2.45% (screams) lower F-measures than the speaker-closed experiment. In the corpus-open experiment, the model correctly classified laughter and screams with F-measures of 85.50% and 73.33%, respectively. Next, logistic regression analysis using feature selection is conducted to identify effective acoustic features for discriminating between laughter and screaming. The results reveal that although the acoustic features that effectively discriminate laughter from screams vary across corpora, the harmonics-to-noise ratio (HNR) can be used to discriminate them in both tested corpora. This applicability of the HNR suggests that the presence of an unvoiced vocal segment in laughter might influence discrimination between laughter and screaming.