Global birdsong embeddings enable superior transfer learning for bioacoustic classification

Area of Science:

• Bioacoustics
• Machine Learning
• Conservation Biology

Background:

• Automated bioacoustic analysis is crucial for monitoring biodiversity and habitats.
• Deep learning models have improved signal classification but require extensive labeled data.
• Many species, especially rare ones, lack sufficient data for robust model training.

Purpose of the Study:

• To investigate the use of feature embeddings from audio classification models for identifying novel bioacoustic classes.
• To evaluate the effectiveness of these embeddings across various taxa, including birds, bats, marine mammals, and amphibians.
• To explore the potential for few-shot transfer learning in bioacoustics.

Main Methods:

• Extracted feature embeddings from pre-trained audio classification models.
• Evaluated embeddings on diverse bioacoustic datasets (bird calls, bat calls, marine mammal calls, amphibian calls).
• Assessed classification performance using limited training data (few-shot learning).

Main Results:

• Embeddings from bird vocalization models yielded higher classification quality than general audio embeddings.
• Successful identification of bioacoustic classes beyond the original training data was demonstrated.
• Few-shot learning was enabled by high-quality feature embeddings, even with minimal data.

Conclusions:

• High-quality feature embeddings from large-scale bird sound classifiers are valuable for few-shot transfer learning in bioacoustics.
• This method offers an efficient way to analyze novel bioacoustic data with limited samples.
• The findings support improved biodiversity monitoring and conservation for data-scarce species.