Marine Mammal Acoustic Species Recognition via Metric Embedding Learning with Audio Spectrogram Transformer

Automated classification of marine mammal vocalizations from passive acoustic monitoring data is essential for ecological research, biodiversity assessment, and the management of anthropogenic impacts on marine ecosystems. Existing deep learning approaches achieve high accuracy on predefined species sets but operate within a closed-set paradigm: they cannot accommodate previously unseen species without complete retraining, which limits their practical deployment in the field. In this work, we propose a metric embedding framework for marine mammal acoustic species classification. Our approach combines an Audio Spectrogram Transformer (AST) backbone, pre-trained on the large scale AudioSet dataset, with a projection head trained using a combined ArcFace and online triplet loss. The model learns a 256-dimensional embedding space in which species similarity is captured by cosine distance, enabling closed-set classification, open-set detection of unseen species, and few-shot learning of novel classes from a limited number of reference recordings. We evaluate our method on the Watkins Marine Mammal Sound Database across 25 known species and 6 species held out for open-set and few-shot evaluation. Averaged over three independent training runs, the model attains 92.4 ± 0.5% closed-set accuracy and a macro F1-score of 0.926 ± 0.005, statistically indistinguishable from a fine-tuned softmax classifier on the same backbone (McNemar test, p = 0.625). The same embedding additionally detects unseen species with an AUROC of 0.891±0.026 under a genuine leave-N-species-out protocol, and discovers new species with a macro F1-score of 0.791 at K = 10 under a leakage-free, validation-tuned threshold. These results demonstrate that metric embedding learning provides a flexible and scalable paradigm for marine mammal acoustic monitoring, supporting the integration of new species into deployed systems without retraining