A machine learning approach for classifying and quantifying acoustic diversity

1. Assessing diversity of discretely varying behavior is a classical ethological problem. In particular, the challenge of calculating an individuals' or species' vocal repertoire size is often an important step in ecological and behavioral studies, but a reproducible and broadly applicable method for accomplishing this task is not currently available. 2. We offer a generalizable method to automate the calculation and quantification of acoustic diversity using an unsupervised random forest framework. We tested our method using natural and synthetic datasets of known repertoire sizes that exhibit standardized variation in common acoustic features as well as in recording quality. We tested two approaches to estimate acoustic diversity using the output from unsupervised random forest analyses: (i) cluster analysis to estimate the number of discrete acoustic signals (e.g., repertoire size) and (ii) an estimation of acoustic area in acoustic feature space, as a proxy for repertoire size. 3. We find that our unsupervised analyses classify acoustic structure with high accuracy. Specifically, both approaches accurately estimate element diversity when repertoire size is small to intermediate (5-20 unique elements). However, for larger datasets (20-100 unique elements), we find that calculating the size of the area occupied in acoustic space is a more reliable proxy for estimating repertoire size. 4. We conclude that our implementation of unsupervised random forest analysis offers a generalizable tool that researchers can apply to classify acoustic structure of diverse datasets. Additionally, output from these analyses can be used to compare the distribution and diversity of signals in acoustic space, creating opportunities to quantify and compare the amount of acoustic variation among individuals, populations, or species in a standardized way. We provide R code and examples to aid researchers interested in using these techniques.