Efficacy of automatic vocalization recognition software for anuran monitoring

Abstract/Summary

Surveys of vocalizations are a widely used method for monitoring anurans, but it can be difficult to coordinate standardized data collection across a large geographic area. Digital automated recording systems (ARS) offer a low-cost method for obtaining samples of anuran vocalizations, but the number of recordings can easily overwhelm human listeners. We tested Song Scope, an automatic vocalization recognition software program for personal computers to determine if this type of machine learning approach is currently a viable solution for anuran monitoring. For three species, Song Scope scanned more than 200 h of recordings in 3-20 h at the settings we chose. The software misidentified true calls (false positive) at rates of 2.7%-15.8% per species and failed to detect calls (false negative) in 45%-51% of recordings. There exists a tradeoff between false positive and false negative errors, which can be adjusted by setting the minimum criteria for the recognition software. Users of this approach should carefully consider their reasons for monitoring and how they intend to use the data before creating a large monitoring network.