Orthoptera Passive Acoustic Monitoring Dataset: Unlabelled and Active-Learning-Labelled UK Field Recordings

Overview This dataset comprises passive acoustic monitoring (PAM) field recordings of Orthoptera (grasshoppers and crickets) and associated soundscapes, collected across Oxfordshire, UK, during summer 2024. It accompanies the paper Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier and was used to train and evaluate the PULSE (Passive acoUstic Latent-Space Encoder) framework. The dataset addresses a key gap in bioacoustic machine learning: the scarcity of open-access field recordings for acoustically identifiable insects, whose calls are typically underrepresented in curated sound libraries. The collection contains approximately 130 GB of recordings. The large majority are unlabelled. A subset has been annotated (strongly annotated with bounding boxes for sound events) with species labels obtained through an active learning process, plus additional randomly selected recordings labelled for validation and testing. Collection Setup Recordings are organised as 15-second units. Recordings were collected across 10 sites in Oxfordshire, UK. Each site was equipped with 2 AudioMoth recorders, giving 20 deployed devices in total; one device failed to record, so usable data is available from 19 recorders. AudioMoths are low-cost acoustic loggers, and the resulting recordings differ from curated library data such as Xeno-canto: alongside the target Orthoptera, they capture numerous additional sound sources including birds, wind, rain, and traffic. Recordings also exhibit the characteristic AudioMoth tonal band, a consistent brighter region between roughly 15–25 kHz visible on spectrograms regardless of environment or recording settings. Recordings were captured at a 96 kHz sample rate. 3 sites are held out for test: CON_2, REG_1, REST_3. Validation is randomly subsampled from 7 remaining training sites. Labelled Subset A portion of the recordings carries expert species annotations. Labels were produced using the Whombat annotation tool, following the label taxonomy of ECOSoundSet: Training annotations (train_labels.csv): selected via active learning using the DIRECT method, which targets the most informative and class-balanced examples for annotation under data imbalance. Validation annotations (val_labels.csv): randomly selected. Test annotations (test_labels.csv): randomly selected, drawn from three sites held out entirely from training to support evaluation of model transferability across sites. Among annotated 15-second clips, 51% contain a single species, 38% contain two species, and the remaining 11% contain more than two species, reflecting the overlapping, near-continuous calling typical of Orthoptera. Four species have substantial presence in the labelled data: Conocephalus fuscus, Leptophyes punctatissima, Pseudochorthippus parallelus, and Roeseliana roeselii (the most frequent). The label taxonomy additionally supports a broader set of UK Orthoptera species and several background categories (Anthropophony, Aves, Bat, Geophony, Passeriformes, Strigiformes). Label Vocabulary (list_of_labels_label_categories.csv) This file lists the complete set of labels and their corresponding label categories that annotators could assign. It contains 386 labels organised into 22 categories, with two columns: label — the specific label (e.g. a species such as Conocephalus fuscus, or a non-biological sound source such as Car). label_category — the higher-level category the label belongs to. The 22 categories span target taxa, other biophony, and non-biological sound. This label taxonomy is taken from the ECOSoundSet dataset: Funosas, D., Massol, E., Bas, Y. et al. A finely annotated dataset for the automated acoustic identification of European Orthoptera and Cicadidae. Sci Data (2026). https://doi.org/10.1038/s41597-026-07150-1 Source record: https://zenodo.org/records/18636037 Ambiguous Labels Annotators were able to express uncertainty about a sound event rather than committing to a single confident species label. This information is preserved in the annotations through two fields: a primary label and an optional ambiguity field (label_ambiguity for mild uncertainty, or label_ambiguity_50 for stronger, roughly even uncertainty). The taxonomy also distinguishes between a specific species label and the generic orthoptera label (used when the annotator is confident the sound is an Orthopteran but cannot resolve the species). To use these annotations for multilabel classification, the labels are converted into soft multihot target vectors. Each annotation type maps to a distribution of probability mass across the class set as follows (where N is the number of Orthoptera classes; mass from confident labels in other bounding boxes takes precedence): label: species — confident species label. The corresponding species receives a value of 1 in the multihot vector. label: species + label_ambiguity: species — mostly confident, with a named alternative species. The primary species receives 0.7; the alternative species receives 0.3. label: species + label_ambiguity: orthoptera — mostly confident, but the alternative is an unspecified Orthopteran rather than a named species. The primary species receives 0.7; the remaining 0.3 is spread uniformly across all other Orthopteran classes as 0.3 / (N − 1). label: species + label_ambiguity_50: species — uncertain between two named species. Both species receive 0.5. label: species + label_ambiguity_50: orthoptera — uncertain, with an unspecified Orthopteran alternative. The primary species receives 0.5; the remaining 0.5 is spread uniformly as 0.5 / (N − 1) across all other classes. label: orthoptera — confident it is Orthoptera, but the species is unknown. Mass is spread uniformly as 1 / N across all classes. label: orthoptera + label_ambiguity: <other class> — mostly confident Orthoptera label. Each class receives 0.7 / N. label: orthoptera + label_ambiguity_50: <other class> — uncertain Orthoptera label. Each class receives 0.5 / N. label: <non-orthoptera class> + label_ambiguity: orthoptera — very uncertain Orthoptera label (the primary call is something else, with Orthoptera as a weak alternative). Each class receives 0.3 / N. This scheme lets downstream models incorporate annotator uncertainty directly rather than discarding ambiguous events or forcing them into hard single-class assignments. Intended Use The unlabelled recordings support self-supervised and domain-adaptation approaches (e.g. learning a local acoustic profile, or distilling from general-purpose bioacoustic models), while the labelled subset supports supervised and weakly supervised Orthoptera classification and benchmarking. The site-based held-out test split enables evaluation of cross-site transferability. Dataset Structure The dataset is uploaded as one ZIP file per site, each named [site_name].zip. Unpacking a site archive yields the following directory structure: [site_name]/ └── AM_[number]/ └── [date]_[time].WAV where: [site_name] identifies one of the 10 Oxfordshire recording sites. AM_[number] is the AudioMoth device identifier. Each site contains up to two such directories, one per deployed device. One device across the deployment failed to record, so its directory is absent for the affected site. [date]_[time].WAV are the individual recordings, named by capture date and time. Licence This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. Associated Paper O. Isupova, D. Kuzin, E. Browning, T. Mills, S. Reece. Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier. ICML 2026 Workshop on Machine Learning for Audio. Paper: https://arxiv.org/abs/2606.13236 Citation If you use this dataset, please cite: O. Isupova, D. Kuzin, E. Browning, T. Mills, S. Reece. Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier. ICML 2026 Workshop on Machine Learning for Audio.