Predicting Concert Hall Reverberation Time and Clarity from Geometric Features Using Machine Learning

Abstract

Concert hall acoustics are traditionally evaluated through direct measurements or acoustic simulations, both of which require significant resources and expertise. This study explores the potential of machine learning (ML) algorithms to predict reverberation time (RT) and clarity (C80) parameters based solely on geometric attributes of concert halls. We analyzed data from 58 concert halls, utilizing 10 geometric features and acoustic measurements across six octave bands (125-4000 Hz). Twenty-eight different ML algorithms were trained and validated using 5-fold cross-validation, with performance evaluated using coefficient of determination (R²) and root mean square error (RMSE). The results demonstrate that geometric attributes explain up to 47% of RT variations at lower frequencies (125-500 Hz) and up to 47% of C80 variations at mid-frequencies (250-1000 Hz). Support vector machines and Gaussian process regression models showed the best performance for RT and C80 estimation, respectively. These findings provide insights into the frequency-dependent influence of architectural geometry on concert hall acoustics and establish a foundation for ML-assisted acoustic design and renovation planning.