InsectSet32: Acoustic Recordings for Insect Species Identification

InsectSet32 contains 335 audio files totaling 57 minutes from 32 sound-producing insect species. The dataset is split between nine Orthoptera species (147 recordings) and 23 Cicadidae species (188 recordings), curated from the unpublished work of Baudewijn Odé and the Global Cicada Sound Collection on Bioacoustica. It was compiled to train neural networks for automatic insect identification, comparing adaptive waveform-based and mel-spectrogram audio frontends.

Use Cases

• Training neural networks for automatic insect species identification based on audio features.
• Comparing the performance of adaptive waveform-based frontends against conventional mel-spectrogram frontends for audio feature extraction.
• Replicating the machine learning experiments from the associated PLOS Computational Biology publication.
• Studying the acoustic signatures of specific Orthoptera and Cicadidae species.

Strengths

• Contains 335 manually inspected audio files, with noisy or multi-species recordings removed to ensure quality.
• Provides annotations for each recording, including species name, identifier, and data subset (training/test) used for neural network training.
• Audio files are standardized as 44.1 kHz mono WAV files, ensuring consistency for machine learning applications.

Limitations

• The dataset scale is limited, with between 4 and 22 files per species, which may constrain model training for some species.
• Metadata completeness is low; specific column names, file sizes, and the exact date of the last update are not provided in the sources.
• A portion of the cicada recordings consist only of the last ten seconds of the original audio files to avoid speech annotations, potentially truncating full acoustic signals.

Provenance

Source

Compiled by Marius Faiß (Leiden University) from the unpublished work of Baudewijn Odé (Orthoptera) and the Global Cicada Sound Collection hosted on Bioacoustica (Cicadidae).

Collection Method

Recordings were manually selected, inspected, and cleaned. Original files were resampled to a consistent 44.1 kHz mono WAV format.