Neural network predictions of acoustical parameters in multi-purpose performance halls

Area of Science:

• Acoustics
• Computational intelligence
• Architectural acoustics

Background:

• Accurate acoustic prediction is crucial for designing multi-purpose performance halls.
• Traditional acoustic measurements can be time-consuming and require extensive data.
• Neural networks offer a potential solution for efficient acoustic property prediction.

Purpose of the Study:

• To evaluate the effectiveness of neural networks in predicting acoustic properties of performance halls.
• To determine the optimal number and placement of measurement points for neural network training.
• To assess the influence of training data volume on prediction accuracy for different acoustic parameters.

Main Methods:

• Binaural sound measurements were conducted in two Hong Kong performance halls of varying capacities and designs.
• A neural network model was developed using audience locations as input.
• Evaluation criteria included root-mean-square deviation, statistical distribution matching, and t-tests comparing simulations and measurements.

Main Results:

• A neural network training scheme with nine uniformly located measurement points per hall area proved most effective.
• Neural network prediction of hall spaciousness achieved accuracy without extensive training data.
• Prediction accuracy for reverberance-related parameters improved with increased training data volume.

Conclusions:

• A neural network model with nine measurement points offers an efficient and effective method for predicting acoustic properties in diverse performance halls.
• The data requirements for neural network acoustic prediction vary by parameter, with spaciousness being less data-intensive than reverberance.
• This approach provides a valuable tool for acoustic design and optimization in architectural acoustics.