Sound quality prediction based on systematic metric selection and shrinkage: Comparison of stepwise, lasso, and elastic-net algorithms and clustering preprocessing

Sound quality is the impression of quality that is transmitted by the sound of a device. Its importance in sound and acoustical design of consumer products no longer needs to be demonstrated. One of the challenges is the creation of a prediction model that is able to predict the results of a listening test while using metrics derived from the sound stimuli. Often, these models are either derived using linear regression on a limited set of experimenter-selected metrics, or using more complex algorithms such as neural networks. In the former case, the user-selected metrics can bias the model and reflect the engineer pre-conceived idea of sound quality while missing potential features. In the latter case, although prediction might be efficient, the model is often in the form of a black-box which is difficult to use as a sound design guideline for engineers. In this paper, preprocessing by participants clustering and three different algorithms are compared in order to construct a sound quality prediction model that does not suffer from these limitations. The lasso, elastic-net and stepwise algorithms are tested for listening tests of consumer product for which 91 metrics are used as potential predictors. Based on the reported results, it is shown that the most promising algorithm is the lasso which is able to (1) efficiently limit the number of metrics, (2) most accurately predict the results of listening tests, and (3) provide a meaningful model that can be used as understandable design guidelines.