Quantifying crowd engagement with machine learning

Inferring a crowd’s engagement at a sporting event using only acoustic signals is difficult for humans to do precisely, but it may be a great job for machine learning. This presentation will summarize the results of using acoustic data from crowds at four types of college sporting events to train a random forest machine-learning model to identify behaviors such as cheering, chanting, and distracting the opposing team. The model uses spectral data and other features derived from acoustic measurements such as spectral slope, spectral kurtosis, and zero cross rate to quantify the level of crowd engagement. This presentation will then discuss an ongoing investigation of how the model performs at predicting crowd engagement from acoustic data collected with a new data-collection system. It will discuss specific challenges and opportunities associated with incorporating new data into an existing machine-learning pipeline.