2025/10/01 by Mitchell C. Cutler, Jason D. Bickmore, Mark K. Transtrum +4 · 1 voice
Computer Science · #Music and Audio Processing #Time Series Analysis and Forecasting #Anomaly Detection Techniques and Applications
paper · doi:10.1121/10.0039709
openalex publication_date 2025/10/01 · openalex created_date 2025/10/24 · openalex updated_date 2026/07/31
Crowds at collegiate basketball games react acoustically to events on the court in many ways, including applauding, chanting, cheering, and making distracting noises. Acoustic features can be extracted from recordings of crowds at basketball games to train machine learning models to classify crowd reactions. Such models may help identify crowd mood, which could help players secure fair contracts, venues refine fan experience, and safety personnel improve emergency response services or to minimize conflict in policing. By exposing the key features in these models, feature selection highlights physical insights about crowd noise, reduces computational costs, and often improves model performance. Feature selection is performed using random forests and least absolute shrinkage and selection operator logistic regression to identify the most useful acoustic features for identifying and classifying crowd reactions. The importance of including short-term feature temporal histories in the feature vector is also evaluated. Features related to specific 1/3-octave band shapes, sound level, and tonality are highly relevant for classifying crowd reactions. Additionally, the inclusion of feature temporal histories can increase classifier accuracies by up to 12%. Interestingly, some features are better predictors of future crowd reactions than current reactions. Reduced feature sets are human-interpretable on a case-by-case basis for the crowd reactions they predict.