Google DeepMind's Bird-Trained Perch 2.0 Model Classifies Whale Calls via Transfer Learning

Google DeepMind's Perch 2.0, a bioacoustics foundation model trained on millions of recordings of birds and other land animals including amphibians, insects and mammals, was the best or second-best model at classifying whale and other underwater sounds across three marine audio datasets, IEEE Spectrum reported in March 2026. The results were presented in a paper at the NeurIPS workshop on AI for Non-Human Animal Communication in December 2025. "If [Perch 2.0] performs well for our whale use cases, then that means we don't need to build an entirely separate new whale model—we can just build on top of that," said Lauren Harrell, a data scientist at Google Research. Google DeepMind and Google Research have studied whale bioacoustics for almost a decade, building algorithms that detect humpback whale calls and a multispecies whale model that identifies eight species and several call types for two of them. Researchers keep finding new call types and unexplained ocean noises, so a single fixed model quickly falls behind. Reusing Perch 2.0 was meant to save computation time and experimentation effort. The team converted each 5-second window of audio into a spectrogram and passed it through Perch 2.0 to produce embeddings, feature sets that capture distinctions such as the difference between humpback and orca whistles. They then randomly picked between 4 and 32 embeddings per dataset to train a logistic regression classifier. The classifier performed well with only a handful of examples and improved as more were added. Perch 2.0 was compared with other bird bioacoustics models, the earlier multispecies whale model and models trained on coral reef sounds; the bird models also did well. The researchers offer three explanations: birds and marine mammals may have evolved similar vocal mechanisms, large models trained on diverse data tend to transfer to out-of-domain tasks, and classifying bird calls forces the model to pick up fine-grained acoustic features that also matter underwater. Killer whale whistles, for example, sit in a similar spectrogram range to many bird vocalizations. Google hopes the model, already used in bird conservation, can support passive acoustic monitoring to protect whale populations.