Data-driven decomposition of crowd noise from indoor sporting events

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Sporting event recordings contain complex acoustic mixtures, making it difficult to isolate specific sound sources like crowd noise.
• Existing methods struggle to effectively differentiate between crowd sounds, music, individual voices, and public address (PA) systems in these environments.

Purpose of the Study:

• To present a data-driven decomposition method for separating diverse acoustic sources from sporting event recordings.
• To analyze the spectral characteristics of sound levels in collegiate sporting events across different sports, crowd sizes, and venues.

Main Methods:

• Utilized machine learning techniques, specifically principal component analysis (PCA), on spectrograms from 30 collegiate sporting events.
• Analyzed 87.5% of spectral variation using three principal components, regardless of sport, venue, or crowd composition.
• Applied Gaussian mixture model clustering to the three-dimensional component coefficient representation to identify distinct acoustic source clusters.

Main Results:

• Identified three principal spectral shapes that effectively separate various acoustic sources.
• Found that three principal components could represent 87.5% of the spectral variation in the audio signals.
• Discovered nine distinct clusters using Gaussian mixture modeling, which audibly separated different combinations of acoustic sources.

Conclusions:

• The proposed data-driven decomposition method, based on PCA and Gaussian mixture modeling, can successfully separate complex acoustic sources in sporting event recordings.
• This approach provides a robust way to differentiate between crowd noise, music, individual voices, and PA systems.
• The findings offer a valuable tool for analyzing and understanding the acoustic landscape of live sporting events.