Google DeepMind's Perch 2.0 foundation model decodes whale calls using bird-trained audio embeddings

Title

Google DeepMind's Perch 2.0 foundation model decodes whale calls using bird-trained audio embeddings

Content

Google DeepMind's Perch 2.0, an AI audio foundation model trained on millions of bird and other animal recordings, achieved best-or-second-best performance on three marine audio datasets for classifying whale vocalizations — using as few as 4 to 32 sample embeddings per dataset to train a logistic regression classifier on top of the foundation model. Lauren Harrell, a data scientist at Google Research, said: "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." Harrell added that transfer learning from a foundation model means researchers can "recycle all of the training that's been done and just do a small model at the end for your use cases," because "there's so many mysterious ocean noises that you can't just have one fixed model." Whale bioacoustics is bottlenecked by both the cost of training species-specific models and an ever-evolving taxonomy of call types — researchers are constantly finding new variants and unknown ocean sounds that fixed, narrow models cannot accommodate. A foundation model already trained on a much larger and more diverse bioacoustic corpus removes the need to retrain from scratch for each new species or call type, dramatically lowering the cost of expanding coverage. Google DeepMind and Google Research have been studying whale bioacoustics for almost a decade, with prior work including CNNs that detect humpback calls and a multispecies whale model that identifies eight distinct species and multiple calls for two of them. The Perch 2.0 work — detailed in a paper (arxiv.org/abs/2512.03219) presented at the December 2025 NeurIPS workshop on AI for Non-Human Animal Communication — takes a foundation-model approach, repurposing a bird-and-land-animal audio model for the underwater domain. The team evaluated Perch 2.0 by converting each 5-second window of audio into a spectrogram, feeding the images through the model to produce embeddings, and then training a logistic regression classifier on a small sample of those embeddings. They compared Perch 2.0 against other bird bioacoustics models, the prior multispecies whale model, and models trained on coral reef sounds, finding that Perch 2.0 was either the best or second-best performer on each comparison. The model is being positioned for passive acoustic monitoring of whale populations — tracking and characterizing whale calls without disturbing marine life — and Google hopes to use it to protect whales while uncovering patterns in their vocalizations. Perch 2.0 continues to serve its original bird conservation applications in parallel, with the new whale use cases building on the same foundation rather than retraining. Per the article, the team plans to keep building on Perch 2.0 for new whale use cases rather than training separate species-specific models, and the broader goal is to use passive acoustic monitoring to "help them unveil the wisdom that these ancient oceanic creatures hold," positioning Perch 2.0 as a foundation for ongoing bioacoustics research across species.