Abstract by Dallen Stark
Audio Feature Reduction of Basketball Crowd Noise
Our group is using machine learning to connect the acoustic signal of sporting event crowds to that crowd's emotional state. My initial work has focused on distinguishing two crowd events from each other (e.g. cheer vs. chant). Recordings of the crowd have been collected at BYU basketball, volleyball, soccer, and football games. Raw data is processed into spectral and low-level features. This feature set is very large, containing over 500 features, so learned models may have large uncertainty in their predictions. I am applying various feature reduction methods to identify the relevant variables for extracting the state of the crowd from the acoustic signal.