Objective-Subjective Sound Quality Correlation Performance Comparison of Genetic Algorithm Based Regression Models and Neural Network Based Approach

Every product is growingly being evaluated in terms of acoustic characteristics. The most accurate way to rate sound quality is by performing jury tests; however, jury tests require a lot of time and human resources. To overcome this problem, jury tests results can be correlated to objective sound quality metrics owing to the fact that objective metrics could be easily obtained from sound data. In this study, advanced techniques for feature identification are explored to correlate objective metrics to subjective perception retrieved from jury tests. The data set referes to the interior noise of a regional propeller aircraft. Artificial Neural Network and two regression models (i.e. linear and quadratic regression models) have been chosen to predict subjective metrics according to the objective data. To obtain the optimized model parameters for the regression models, a Genetic Algorithm has been used as optimization strategy. In each modelling, 85 percent of sound sample data have been utilized to perform the model and remaining 15 percent have been reserved for testing the models. The results showed that the Artificial Neural Network can provide better prediction.