ABSTRACT Killer whales use complex vocalizations to maintain social structure and coordinate behavior, yet automated classification of call types remains challenging due to overlapping calls and structural similarity among call types. I evaluated the performance of the deep learning framework ANIMAL‐SPOT for identifying killer whale call types from raw acoustic recordings. Using a training set of stereotyped calls recorded from killer whales in Avacha Gulf, Russia, the ANIMAL‐SPOT model was trained with two window‐step parameter combinations. Detection across the raw recordings was generally successful for high‐ and moderate‐quality calls. For the call type identification, the main issue was excessive segmentation, where a single call was mistakenly classified as 2–3 different call types due to structural similarities among certain syllables in different call types. Overall classification accuracy averaged 47.8%, with errors reflecting structural similarities among call types and syllables. I discuss improvements including syllable‐level training, two‐stage detection pipelines, and probability‐based annotation adjustments. Importantly, misclassification patterns provide insights into acoustic similarity, offering a novel approach to studying call structure and cultural evolution. While current results are limited by segmentation challenges, this study demonstrates both the potential and the constraints of deep learning for automating killer whale vocal repertoire analysis.
Olga A. Filatova (Mon,) studied this question.