Deep Learning for Orca Call Type Identification - A Fully Unsupervised Approach

Call type classification is an important instrument in bioacoustic research investigating group-specific vocal repertoire, behavioral patterns, and cultures of different animal groups. There is a growing need using robust machine-based techniques to replace human classification due to its advantages in handling large datasets, delivering consistent results, removing perceptualbased classification, and minimizing human errors. The current work is the first adopting a two-stage fully unsupervised approach on previous machine-segmented orca data to identify orca sound types using deep learning together with one of the largest bioacoustic datasets – the Orchive. The proposed methods include: (1) unsupervised feature learning using an undercomplete ResNet18-autoencoder trained on machineannotated data, and (2) spectral clustering utilizing compressed orca feature representations. An existing human-labeled orca dataset was clustered, including 514 signals distributed over 12 classes. This two-stage fully unsupervised approach is an initial study to (1) examine machine-generated clusters against human-identified orca call type classes, (2) compare supervised call type classification versus unsupervised call type clustering, and (3) verify the general feasibility of a completely unsupervised approach based on machine-labeled orca data resulting in a major progress within the research field of animal linguistics, by deriving a much deeper understanding and facilitating totally new insights and opportunities.