The use of artificial neural networks to predict tonal sound annoyance based on noise metrics and psychoacoustics parameters

Abstract

The tonal noise of indoor mechanical systems causes an unpleasant sensation. The present study was conducted to predict tonal sound annoyance based on noise metrics and psychoacoustics parameters using artificial neural networks. Thirty-six signals of noise were produced by six tone levels, three tone frequencies, and two background noise levels in an enclosed space. Then, noise metrics and psychoacoustic parameters of the signals were determined. Subsequently, 60 subjects were asked to express their subjective perception of annoyance during exposure to various noises. Finally, the predictive model of annoyance was computed using the feed-forward neural networks. The initialization of weights and biases was performed using the Nguyen-Widrow method. The gradient descent with momentum and back-propagation algorithms were applied to learn the function and network weights, respectively. Based on the results, higher tone level, higher background noise level, lower frequency, and less sharp noise significantly increased the value of the perceived annoyance. The obtained Kaiser-Mayer-Olkin coefficient of the model was equal to 0.8. The values of recognition rate related to data of training and testing were computed by 0.83 and 0.91, respectively. The parameters of loudness, audible tone, and roughness compared to combined metrics were more substantial predictors of perceived annoyance.