Colloquium: Vlad Stefan Buzetelu (FPT)

Concerns regarding the effects of aircraft noise on the health and well-being of communities living in the vicinity of airports have been increasing, and there is a lack of consensus in the scientific community on which metrics are the best predictors for this type of annoyance. This study aims at developing a machine learning methodology for instant sound metric predictions and for annoyance rating predictions from flyover recordings. The framework involves a Convolutional Neural Network (CNN) for the former, followed by models which use the CNN predictions as input for the latter. A campaign of listening experiments was conducted to gather annoyance data from 60 aircraft flyover recordings, and a correlation analysis was subsequently made on a large pool of metrics, to isolate those with the largest predictive potential. In general, metrics derived from Psychoacoustic Annoyance models present better performance as predictors compared to conventional metrics and most Sound Quality Metrics taken individually. The AI framework achieves very promising results in terms of errors for both the annoyance ratings and metric predictions.

Supervisor: Roberto Merino-Martinez