“Autonomous sensors and machine learning are transforming ecology by enabling large-scale observation of organisms and ecosystems. However, camera traps, audio recorders and other sensors currently produce species-level detections. Tracking individuals continues to require invasive, high-effort capture–mark–recapture. Here we introduce an automated, general approach that identifies individual animals in passive acoustic recordings based on individually distinctive vocalizations. We train a convolutional neural network to extract salient features from audio recordings using a location-prediction task, then apply dimensionality reduction and clustering on features to achieve song-based individual identification. We identify individuals not included in the training data with 96% accuracy.”

From Nature.