Learning to detect an animal sound from five examples

Automatic detection and classification of animal sounds has many applications in biodiversity monitoring behavior. In the past twenty years, volume digitised wildlife sound available massively increased, automatic through deep learning now shows strong results. However, bioacoustics is not a single task but vast range small-scale tasks (such as individual ID, call type, emotional indication) with wide variety data characteristics, most bioacoustic do come strongly-labelled training data. The standard paradigm supervised learning, focussed on large-scale dataset and/or generic pre-trained algorithm, insufficient. this work we recast event within AI framework few-shot learning. We adapt to detection, such that system can be given annotated start/end times few 5 events, then detect events long-duration audio—even when category was known at time algorithm training. introduce collection open datasets designed strongly test system's ability perform detections, present results public contest address task. Our analysis prototypical networks are very common used strategy they well enhanced adaptations for general characteristics sounds. systems high resolution capabilities best challenge. demonstrate widely-varying durations an important factor performance, non-stationarity, i.e. gradual changes conditions throughout duration recording. For fine-grained recognition without massive data, our powerful new method, outperforming traditional signal-processing methods fully automated scenario.