InsectSet47 & InsectSet66: Audio Datasets for Insect Species Identification
by Marius Faiß / Leiden University
Available on 1 platform
Sign in to view source links and access this dataset
Description
Two audio datasets compiled for training neural networks to automatically identify insect species from their sounds. InsectSet47 contains 1,006 recordings from 47 species totaling 22 hours, while InsectSet66 expands to 1,554 recordings from 66 species totaling over 24 hours. The datasets were created by Marius Faiß at Leiden University, sourced from BioAcoustica, xeno-canto, iNaturalist, and private collections.
Use Cases
Training neural networks for insect species identification based on standardized audio recordings.
Comparing adaptive waveform-based frontends to mel-spectrogram frontends for audio feature extraction.
Evaluating model performance on citizen-science sourced audio data with a 60/20/20 train/validation/test split.
Strengths
Contains over 24 hours of audio across 66 species in InsectSet66.
Ensures a minimum of ten audio files per species for model training.
Files were manually inspected and cleaned to remove noise interference and multi-species recordings.
Dataset splits (60/20/20 by file count) were designed to prevent data leakage.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Data may reflect geographic or taxonomic bias inherent to the contributing citizen science platforms.
Row count and specific file formats beyond WAV are unknown.
Provenance
Source
BioAcoustica, xeno-canto, iNaturalist, and private collections by Baudewijn Odé.
Collection Method
Recordings were downloaded, manually inspected, and standardized to 44.1 kHz mono WAV files.
Geography
Recordings sourced from citizen scientists worldwide.
Annotation files mark recordings edited into smaller files; these should be assigned together into train/validation/test sets to prevent data leakage.