Automated parsing of sound level meter data into artifacts, natural sources, and noise

Abstract

The National Park Service (NPS) has nearly one million hours of 1/3rd octave band sound level measurements collected at hundreds of sites throughout the system. One of the principal purposes of these data is to measure the background or residual sound level against which all transient sounds are heard. Historically this analysis has depended upon trained listeners to identify which sound sources can be heard in segments of these data. These annotated records were then used to calculate an adjusted median sound level. To pursue a more efficient, automated method to deliver this result, matrix decomposition methods were tested to characterize their capacity to model time series of sound level spectra as low-rank decompositions, including options to account for anomalous events. These methods were combined with sound source identification based on component spectral properties and time weightings to yield promising approaches for automating NPS analyses. These results may also offer options for detecting and removing the effects of pseudonoise due to the turbulence generated by air flowing past the microphone.