TRUST, BUT VERIFY: PITFALLS AND SOLUTIONS FOR VERIFYING AUTOMATED ACOUSTIC CLASSIFIER OUTPUT

Jerry S Cole; The Institute for Bird Populations; jcole@birdpop.org; Mary K. Clapp, Katelyn A. Ray, Joe Weiss

Tools for processing acoustic recordings of wildlife have rapidly advanced in recent years. The advent of automated classifiers like BirdNET and Perch have enabled researchers to identify thousands of species in large quantities of audio. However, methods and standards for summarizing and manually verifying the output from these classifiers are still being actively developed. We provide a brief overview of potential pitfalls ecologists may encounter when verifying and analyzing automated sound classifier output using existing methods, and use simulated acoustic classifier output to demonstrate the consequences of sub-optimal verification procedures. We also present preliminary methods for reducing verification effort and tests to determine when additional verification effort is needed to reduce bias in downstream analyses of discrete data sets.

Advances in Avian Research and Monitoring 
Friday 9:45 AM
 

Speaker Bio:

Jerry has worked on various avian projects throughout the West, ranging from tracking Clark's Nutcrackers to helping capture California Condors in southern California. Jerry currently works on analyzing in-person and automated acoustic bird survey data collected within California, Colorado, and other western states. He is also interested in contributing to the development of methods to turn acoustic classifier output into ecological insights.