Optimal pretraining strategies for few-shot bioacoustic transfer learning

Determine the optimal pretraining strategies to produce pretrained networks that enable accurate few-shot transfer learning under substantial domain shift and limited annotated data in novel bioacoustic domains.

Background

The paper targets transfer learning for passive acoustic monitoring in data-deficient marine environments, especially coral reefs, where annotated datasets are scarce and domain shift from terrestrial datasets is substantial.

While pretrained models exist for birds and general audio, the authors note that crafting novel pretrained networks tailored for few-shot learning in new bioacoustic domains is hampered by uncertainty about which pretraining strategies are best, particularly given limited labeled data.

References

Substantial domain shifts may require the development of novel pretrained networks to achieve accurate few-shot transfer learning. The optimal pretraining strategies to produce these networks remain unknown, which is further compounded by the sparsity of annotated libraries for novel domains.

— Leveraging tropical reef, bird and unrelated sounds for superior transfer learning in marine bioacoustics  (2404.16436 - Williams et al., 2024) in Introduction

While fine-tuning with dolphin-specific data mitigates this concern, the optimal architecture for cetacean vocalization analysis remains an open question.

— Interpreting Dolphin Vocal Sequences via Multiple Sequence Alignment  (2609.08795 - Kohlsdorf et al., 8 Sep 2026) in Discussion, paragraph beginning “Several limitations warrant consideration.”