Sound Quality Prediction Using Neural Networks

Abstract: Sound quality is an important measure depicting the quality of a machine as well as the comfort in its usage. However, it being a subjective measure, not only is it difficult to capture it ahead of time but also necessitates time and cost expensive jury testing. Thus, it is worthwhile to be able to effectively predict the results of the jury study from metrics that can be objectively measured. The aim of the thesis is twofold: first, to establish neural network models to predict subjective sound quality metrics from objective metrics and second, to interpret the model to understand the relative importance of each objective metric towards a specific subjective rating. Ultimately the thesis aims to pave the way for inclusion of sound quality metrics in the early design stages. From the study, it was evident that neural networks’ performance was at least equal to or better than linear or quadratic models. The connection weights method, the profile method, the perturbation method, the improved stepwise selection method and linear regression method were the interpretation algorithms found to work well in all simulated data-sets. They also gave comparable results on the real data-sets. Neural networks were shown to have the potential for giving low prediction errors while maintaining interpretability in sound quality applications. The data scarcity study gave an idea of the scale of performance enhancement that can be achieved with more data and can act as a useful input for deciding the number data points.