We must not fool ourselves: A reply to Sethi et al. on the use of BirdNET to classify neotropical birdcalls

Abstract

The acousticmonitoring (PAM) of biodiversity has gained traction in recent years, even though classifying species within a recording could be challenging in places where acoustic diversity is high. Among the classification algorithms recently developed, BirdNET is probably the most promising. BirdNET was built to recognize over six thousand bird species and was trained using data from Xeno-canto and the Macaulay Library. Despite its huge potential, BirdNET is known to struggle with noisier recordings (1), reducing its accuracy for PAM.