Principal component analysis applied to feature selection at collegiate basketball and football games

A supervised machine learning model trained acoustic features has been used to previously identify crowd noise reactions during collegiate basketball games. Three of these features come from principal component analysis on one-third octave band spectra [M. C. Cutler et al., J. Acoust. Soc. Am. 155, 962–970 (2024)]. Principal component analysis was repeated on a new dataset of crowd noise from college football games. Despite differences in venue, crowd size, and reactions dependent on event type, the principal component vectors were similar. In both cases, the top three principal components relate to overall sound pressure level, distinguishing crowd noise from public address (PA) noise, and distinguishing crowd noise from music played over the PA system. The results suggest the viability of this approach in identifying crowd reactions in recordings with other noise sources present.