Pilot study of perception of wind farm noise classified by cluster analysis

From a publicly-available database of 3000 10-second samples of wind farm noise (WFN) (https://open.flinders.edu.au/articles/dataset/Benchmarkwind_farm_noise_data_set/19618023/1), we extracted 840 samples that had no appreciable environmental sounds using YAMNet analysis, spectrogram inspection, and listening. The 31 parameters that have been used to identify amplitude modulation (AM) via machine learning (https://github.com/ducphucnguyen/WFNAM_Detection) were calculated for the 840 identified samples. The resulting values were used to group the samples via cluster analysis. This separated the files into clusters with qualitative differences, e.g., broad spectrum with little or no apparent amplitude modulation (AM), broad spectrum with apparent AM, predominantly low-frequency spectrum with pulsating low-frequency tones, etc. We present the results of listening tests on noise samples that are typical of the different clusters. Samples were rated by young adults with no prior listening training who passed a hearing screening. Testing was conducted in a single-walled audiometric test suite. Test stimuli were normalized to have equal loudness in sones as computed per ISO532-1. Participants rated the loudness and unpleasantness of the stimuli via magnitude estimation and a paired comparison method. Results will be presented and discussed in view of the classification of the wind farm noise stimuli by cluster analysis.