Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier

The article presents PULSE, a semi-supervised, multi-task framework for classifying Orthoptera bioacoustics, which integrates weakly-supervised species classification, self-supervised learning from unlabelled audio, and knowledge distillation. PULSE significantly outperforms a state-of-the-art general model with macro F1 scores of 0.21 versus 0.07 and AUC of 0.74 versus 0.45, with active learning improving these metrics to 0.34 and 0.84, respectively. This framework enhances ecological monitoring capabilities by providing meaningful embeddings and an interactive visualization tool, making it a valuable resource for practitioners in bioacoustic analysis and ecological inference.