Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Life Sciences & Biology, Environmental Science & Earth Systems · Depth: Expert, extended

Summary

PULSE, a novel semi-supervised, multi-task framework, significantly advances Orthoptera bioacoustic classification for ecological inference. Designed to overcome the narrow training and non-transferability of existing passive acoustic monitoring tools, PULSE integrates weakly-supervised species classification, self-supervised learning on unlabelled field audio, and knowledge distillation from the BirdNET model. The framework was trained using publicly available labelled Orthoptera data and nearly 150 GB of unlabelled UK field recordings from 10 Oxfordshire sites. PULSE demonstrated superior performance over a leading general bioacoustic model, Perch 2.0, achieving a macro F1 of 0.21 versus 0.07, AUC of 0.74 versus 0.45, and AP of 0.32 versus 0.19. Active learning further boosted its F1 to 0.34 and AUC to 0.84. Beyond classification, PULSE's learned embeddings reveal ecologically meaningful structures and enable the unmixing of overlapping calls, providing vital information for conservation and habitat management.

Key takeaway

For Machine Learning Engineers developing bioacoustic classifiers in data-scarce environments, you should adopt multi-task, semi-supervised frameworks. This approach, exemplified by PULSE, significantly improves model performance and transferability compared to purely supervised methods. Integrate knowledge distillation from general models and self-supervised learning on unlabelled data to enhance adaptability. Additionally, consider active learning strategies to efficiently expand your labelled datasets, reducing manual annotation burden while boosting classification accuracy for ecological monitoring.

Key insights

Combining semi-supervised, multi-task learning, and knowledge distillation improves bioacoustic classification with limited labels.

Principles

• Unsupervised tasks are crucial for scarce labelled bioacoustic data.
• Multi-task training mitigates data scarcity and domain shift.
• Distill general audio representations for downstream tasks.

Method

PULSE uses a VGGish backbone on mel spectrograms, optimizing three objectives: supervised classification with binary cross-entropy, BirdNET embedding matching via L2 loss, and BYOL for self-supervision. Losses are jointly optimized, and active learning labels field data.

In practice

• Utilize VGGish backbones for audio spectrogram processing.
• Implement time and frequency masking for spectrogram augmentation.
• Apply NNLS on prototype vectors to unmix multi-species audio.

Topics

• Multitask Learning
• Semisupervised Learning
• Orthoptera Classification
• Bioacoustic Monitoring
• Knowledge Distillation
• Active Learning
• Ecological Inference

Code references

• mbsantiago/whombat

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier — PULSE combines semi-supervised, multi-task learning with knowledge distillation to create a high-performing, domain-adapted bioacoustic classifier for Orthoptera.
• A frequency analysis of filterbank initialisation and noise augmentation for LEAF — LEAF filterbanks show high inertia, resisting adaptation and converging to a fixed S-curve regardless of input frequency content.
• A finely annotated dataset for the automated acoustic identification of European Orthoptera and Cicadidae — A new dataset combines weak and strong acoustic labels for automated European insect identification.
• AI Trained on Birdsong Can Recognize Whale Calls — Google DeepMind's Perch 2.0, an AI audio model primarily trained on millions of land-based animal recordings including birds, amphibians, insects, and mammals, has shown unexpected strong performance…
• LoSATok: Low-dimensional Semantic-Acoustic Tokenizer for Cross-Domain Audio Understanding and Generation — LoSATok unifies audio understanding and generation via low-dimensional semantic-acoustic tokens, reducing Diffusion Transformer modeling complexity.
• GetNetUPAM: Ecologically Informed Nested Cross-Validation and Noise-Robust Attention for Marine Bioacoustic Monitoring — Ecologically informed nested cross-validation and adaptive attention improve marine bioacoustic model stability and generalization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.