Automatic classification of insect sounds

Acoustic monitoring provides a non-invasive approach for studying insect activity and species composition, but automatic insect classification remains challenging due to heterogeneous recording conditions, acoustic similarity between species, and strong class imbalance. In this work, we report experiments conducted within the BioMonitor4CAP EU Horizon Europe project on automatic insect species classification using a heterogeneous bioacoustic dataset, with a strong focus on field and laboratory recordings from Bulgaria.The dataset comprises high-resolution insect audio recordings collected from multiple sources, including recordings of targeted Orthoptera from Bulgaria, recordings used in the book The Grasshoppers of Greece, and publicly available xeno-canto recordings, covering a wide range of insect species. The recordings exhibit substantial variability in duration, sampling rate, and acoustic conditions. Audio signals were resampled and segmented into short, overlapping time–frequency patches, and patch-level predictions were aggregated for event-level inference. For classification, several deep learning architectures were evaluated, with the Patchout Spectrogram Transformer (PaSST) showing the best performance.The results demonstrate the potential of transformer-based models for insect acoustic classification, while also highlighting persistent challenges related to class imbalance, acoustic similarity between species, and fragmented song structure. These findings provide a foundation for future improvements in robust, large-scale insect bioacoustic monitoring.