Automatic Deep-Learning-Based Classification of Bottlenose Dolphin Signature Whistles

Bottlenose dolphins (Tursiops truncatus) produce individually distinctive signature whistles that develop early in life and are used to recognize and maintain contact with conspecifics. Health assessments in Sarasota, Florida (USA), have provided a unique opportunity to record signature whistles of wild individuals of known age, sex, and matrilineal relationships. After 37 years, the Sarasota Dolphin recording library contains 930 recording sessions of 296 individual dolphins. Here, a deep convolutional neural network classifier was trained using a curated subset of 200 different signature whistles from each of 70 individual bottlenose dolphins. A MobileNetV2 trained on spectrogram data allocated signature whistles to the correct individual with an accuracy of 95.8%. To improve generalization to novel audio datasets, a data augmentation step was implemented that incorporated time stretching, pitch shifting, and mixing database whistles with natural background noise recordings. This augmented model achieved a classification accuracy of 94.5% for high signal-to-noise ratio (SNR) signature whistles, and 92.6% under more realistic conditions with signature whistles mixed with ambient noise at -6 to 12 dB SNR. These initial results are promising and suggest that automatic signature whistle classification techniques could enable acoustic monitoring of movements and habitat use of individual bottlenose dolphins at scale.