InterSpeech 2021

Robust Laughter Detection in Noisy Environments
(3 minutes introduction)

Jon Gillick (University of California at Berkeley, USA), Wesley Deng (University of California at Berkeley, USA), Kimiko Ryokai (University of California at Berkeley, USA), David Bamman (University of California at Berkeley, USA)
We investigate the problem of automatically identifying and extracting laughter from audio files in noisy environments. We conduct an empirical evaluation of several machine learning models using audio data of varying sound quality, finding that while previously published methods work relatively well in controlled environments, performance drops precipitously in real-world settings with background noise. In the process, we contribute a new dataset of laughter annotations on top of the existing AudioSet corpus, with precise segmentations for the start and end points of each laugh, and we present a new approach to laughter detection that performs comparatively well in uncontrolled environments. We discuss the utility of our approach as well as the importance of understanding the variability of model performance in a range of real-world testing environments.