Abstract by Spencer Wadsworth
Physics and Astronomy
Machine Learning Classifying of Crowd Acoustics from College Basketball Games
While ML has been applied in numerous audio applications, the aim is usually to distinguish events from noise, rather than trying to characterize the noise itself. This presentation comprises an initial study using ML to characterize crowd dynamics during collegiate basketball games. High-fidelity crowd noise recordings from several men’s and women’s games were synchronized with game video and used to produce a training dataset for supervised ML by linking game events (e.g., baskets, fouls) with acoustic labels (e.g., cheering, silence, applause). Using the training dataset, a ML classifier, specifically a random forest classifier, was built and customized to identify causal game events from acoustic crowd responses. The initial findings are very promising having high levels of accuracy. These findings are discussed, along with potential improvements from adjusting training data and ML classifier selection.