In a surprising demonstration of machine learning's potential, Google Deepmind has unveiled a bioacoustic model that excels in identifying whale songs despite being primarily trained on bird calls. This breakthrough underscores the power of generalization in artificial intelligence, showing how models can transfer skills across vastly different domains.
The research highlights that the model, known as Perch 2.0, outperformed specialized whale-detection models, including Google's own Multispecies Whale Model. This success is attributed to the fine-grained distinctions required for bird call classification, which seem to translate effectively to identifying marine mammal sounds due to similar evolutionary sound production mechanisms shared between birds and marine mammals.
The Perch 2.0 model, consisting of 101.8 million parameters, was trained on over 1.5 million recordings from a diverse range of animal species, predominantly birds. Despite the scarcity of aquatic recordings in its training dataset, the model's performance in marine sound classification tasks is remarkable. It was tested against three datasets, including NOAA PIPAN, ReefSet, and DCLDE 2026, demonstrating exceptional accuracy in these challenging environments.
When tasked with distinguishing orca sounds from different subpopulations, Perch 2.0 achieved an AUC-ROC score of 0.945, significantly outpacing the whale-specific model's score of 0.821. The general-purpose model's ability to classify marine sounds with minimal training data is a testament to its robust design.
The research team identified several reasons for this unexpected cross-domain success. Firstly, larger models tend to generalize better, benefiting from neural scaling laws. Secondly, bird classification demands the detection of minimal acoustic differences, honing the model's ability to discern subtle variations. Lastly, the shared sound production methods between birds and marine mammals enhance the model's ability to transfer knowledge across species.
This advancement holds practical implications for marine bioacoustics, where new sounds are frequently discovered. The model's ability to quickly train classifiers for newly identified sounds could revolutionize the field, allowing for rapid analysis and classification of marine noises.
Google's initiative includes providing open access to the tools and resources necessary for using this model, with tutorials available on Google Colab and code hosted on GitHub. This makes it easier for researchers and enthusiasts to leverage the model's capabilities for further exploration in bioacoustics.