EarthChirp: a global reference library for insect acoustic recognition and discovery across the audible and ultrasonic spectrum

marko melnick 

1. Passive acoustic monitoring (PAM) is scaling rapidly, but automated recognition for insects lags far behind birds. The dominant recogniser (BirdNET) classifies only ~35 insect species and building bespoke insect classifiers requires labelled training data that does not exist for most taxa.

2. We present EarthChirp, a training-free recogniser for singing insects (Orthoptera and Cicadidae) built on a frozen bioacoustic foundation model (Perch 2.0). EarthChirp pairs a global reference library of 2,012 species (assembled from GBIF, Xeno-canto, SINA and InsectSingers) with a single confidence-gated recogniser that returns a named species where the match is strong and an unnamed morphospecies placed phylogenetically against the reference library (reliably only for the dominant group) where it is not.

3. We find detection and discovery generalise zero-shot to independent data, and show EarthChirp separates insects from vertebrate biophony (AUROC 95%) and recovers morphospecies on an independent 104-species European corpus. Fine species naming does not transfer zero-shot across recording conditions (~26% open-set), but saturates near 58% with ten local reference clips per species (recording-disjoint). The residual false-positive risk is anthropogenic mechanical noise, not other animals. Using an outgroup prototype we are able to lift insect-versus-noise discrimination from 74% to 93%.

4. EarthChirp is released as a ~29 MB library and classifier head usable with the public Perch model and requires no GPU and no retraining. It detects and discovers insects in any soundscape and adapts to local faunas with a handful of reference clips, which is a deployment recipe rather than a per-site retraining cost. The same frozen-embedding approach reaches beyond the model's 16 kHz ceiling by time-expansion, recovering ultrasonic insects and bats. On the limited open bat audio currently available, we show EarthChirp generalizes to this unrelated ultrasonic taxon (bats) and provisionally places species absent from its reference library. We map the method's capabilities and limits to guide its use in conservation monitoring.