An Ensemble Deep-Learning System for Few-Shot Bioacoustic Event Detection

Human are good at identifying new objects from only a few number of samples. Inspired by the rapid learning ability of humans, few-shot bioacoustic event detection task based on machine learning method has attracted more attention lately. In this paper, we formulated an ensemble deep-learning system to address the task, incorporating an enhanced prototypical network and transductive inference methodology. We have submitted our models to Task 5 of the Detection and Classification of Acoustic Scenes and Events 2022 (DCASE2022) challenge, and got F-measure of 64.8% on the validation set as our best score. In the final competition, our system got F-measure of 44.3% on the evaluation set.