Assessment of learning algorithms to model perception of sound

Abstract

Predicting human response to complex sound is a nontrivial task. Besides large differences among subjects and practically infinite types of stimuli, human response to sound is typically quantified by a few parameters having nonlinear behavior. Still, such predictions are valuable for the assessment of sound quality, which is a critical step toward the development of systems that offer improved human comfort, productivity and wellness. The overall objective of this work is to learn about human perception of sound through the use of machine learning algorithms. Learning algorithms are ideal for modeling the complex behavior of subjective parameters and identifying new trends with the potential to accumulate knowledge from different experiments. This work compares the performance of four learning algorithms (linear regression, support vector machines, decision trees, and random forests) to predict annoyance due to complex sound. Construction of these models relies on the annoyance response of 38 subjects to 103 sounds described by five known predictors (loudness, roughness, sharpness, tonality, and fluctuation strength). Comparison of these algorithms in terms of prediction accuracy, model interpretability, versatility, and computation time indicates that decision trees and random forests are the best algorithms for this task.