The lasso and elastic-net algorithms for predictive sound quality models using large pool of predictive metrics and factors

Résumé

Part of the overall perceived quality is communicated by sound quality. Sound quality studies present a twofold challenge. The first is that they rely on listening tests. However, listening tests are considered time consuming. The second challenge relates to the development of predictive sound quality models. Often, these models are derived using linear regression on a limited set of predictive metrics. However, nowadays, many potential metrics are available and there is no computational burden that should limit the number of potential metrics. An issue is that regression using more metrics than observations cannot lead to meaningful predictive models since all the metrics are selected. In this paper, different algorithms are compared to construct a sound quality predictive model that does not suffer from these limitations. These algorithms achieve an automatic selection of few metrics from a large pool. The lasso, elastic-net and stepwise algorithms are tested for the prediction of listening tests results of consumer product for which more than 100 metrics are used a potential predictors. It is shown that the most promising algorithm is the lasso which is able to efficiently limit the number of metrics and provide a model that can be used as understandable design guidelines.