Identifying Dolphin Whistle Producers With Deep Learning: Moving Beyond Signature Whistles

Résumé fourni par la source

ABSTRACT Bottlenose dolphins produce several types of whistle contours, including signature whistles, shared whistles, copies of conspecifics' signature whistles, and variant whistles. While signature whistles as individual identifiers are well studied, less has been demonstrated for identifying dolphins from non‐signature whistle types. Shared whistles, which are produced by multiple dolphins, and signature whistle copies are particularly difficult to classify. Variant whistles remain largely unexplored for individual identification. This study evaluated the use of a convolutional neural network operating on raw amplitude vectors to classify which dolphin produced the whistle (20 different dolphins), regardless of the contour type, from the raw time‐series of whistle recordings. When trained on all whistle types, the model achieved strong performance in classifying whistle producers (85.20%), with the highest accuracies for signature (96.10%) and shared whistles (92.50%), moderate accuracy for variant whistles (72.10%), and lowest accuracy for signature whistle copies (17.20%). To evaluate the trade‐off between classification accuracy and ecological relevance, a model run solely on signature whistles achieved 93.70% accuracy. This study represents a first step toward caller identification across whistle types for well‐studied populations, rather than solely relying on individually unique signature whistles alone. These results also demonstrate the potential for deep learning (DL) and raw audio time series to support future advances in passive acoustic monitoring of odontocetes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.