Predicting real indoor soundscapes based on auditory and non-auditory factors across different loudness ranges with linear and nonlinear models

Studies on the indoor sound environment have mainly focused on the impact of traffic, neighbor, or ventilation noise on health, annoyance, or sound quality. When conducting such studies in the laboratory, acoustic stimuli are often tailored to the research question, which presumably do not reflect the variation in the acoustic environment at home. Thus, we re-analyzed soundscape data gathered by participants at their homes several times a day. Since the recorded soundscapes differed notably between participants (and their dwellings) in terms of loudness, we investigated if cross-validated linear regression and Random Forest models predicting the soundscape dimensions based on four loudness quartile subsets achieve higher prediction performance compared to models using the full volume range. Our results, which are based on the explanation of the variance in the targets, the correlation between the targets and acoustic and perceptual predictors, and the predictor importance, suggest the use of Random Forest models based on the entire loudness range. In our analysis of predictor importance, Pleasantness was understandably dominated by perceptual measures, whereas Eventfulness was driven through acoustic and psychoacoustic predictors. Interestingly, all models based on quiet soundscapes behaved differently, which provides scope for investigating the underlying causes in future.