Quantitively measuring differences between ambient sound environments using statistical distance metrics

A recurring challenge in ambient sound studies is determining whether sensors placed at two different locations (e.g., drifting over a seamount versus abyssal plain) or at different times (e.g., before and after a coral reef degrades) experience the same ambient acoustic environment, or “soundscape.” Here, we examine various candidate metrics for quantitively measuring the difference between two ambient environments. The ambient environment is modeled as a multidimensional probability distribution, with different measurements (sound pressure level, dominant directionality) defining different axes. Temporal patterns (such as periodic variations in intensity or impulse sequences) are mapped as additional axes representing time-delayed measurements. Various metrics and divergences like the inner product, relative entropy (Kullback–Leibler divergence), total variation distance, and Wasserstein are then used to compute the “distance” between a given environment and a “reference” environment. The resulting outputs reduce a multidimensional distribution into a single scalar, easing visualization of changes in broadband (potentially directional) ambient sound data. Examples of applying this approach to bulk data are shown on data collected from both deep-water drifters and shallow coral reefs, including an example that helped flag a sensor noise issue. [Work supported by ONR TFO.]