Perch 2.0 transfers'whale'to underwater tasks

2025-12-02

AI Summary

This study investigates the few-shot transferability of the large-scale terrestrial bioacoustic pre-trained model Perch 2.0 to marine mammal and underwater audio classification—despite its training corpus containing virtually no marine species. To address the absence of labeled marine mammal data, we employ linear probing: fine-tuning only the final classification layer using ≤16 labeled underwater audio samples per class. Our experiments constitute the first systematic evaluation of Perch 2.0’s cross-domain, cross-medium (terrestrial → aquatic) generalization capability. Results demonstrate that Perch 2.0 significantly outperforms established bioacoustic models—including PANNs and BirdNET—across multiple marine mammal classification benchmarks. These findings establish Perch 2.0 as a highly effective foundation model for few-shot underwater acoustic classification, enabling high-accuracy marine biodiversity monitoring under extreme label scarcity. The work introduces a novel paradigm for leveraging terrestrial pre-training to advance passive acoustic monitoring in marine ecosystems.

Technology Category

• Machine Learning: Bio-inspired Learning
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Web Mining and Content Analysis: Large pretrained models with web data
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search

Abstract

Perch 2.0 is a supervised bioacoustics foundation model pretrained on 14,597 species, including birds, mammals, amphibians, and insects, and has state-of-the-art performance on multiple benchmarks. Given that Perch 2.0 includes almost no marine mammal audio or classes in the training data, we evaluate Perch 2.0 performance on marine mammal and underwater audio tasks through few-shot transfer learning. We perform linear probing with the embeddings generated from this foundation model and compare performance to other pretrained bioacoustics models. In particular, we compare Perch 2.0 with previous multispecies whale, Perch 1.0, SurfPerch, AVES-bio, BirdAVES, and Birdnet V2.3 models, which have open-source tools for transfer-learning and agile modeling. We show that the embeddings from the Perch 2.0 model have consistently high performance for few-shot transfer learning, generally outperforming alternative embedding models on the majority of tasks, and thus is recommended when developing new linear classifiers for marine mammal classification with few labeled examples.

Problem

• Evaluates Perch 2.0's performance on marine mammal audio tasks
• Compares Perch 2.0 with other bioacoustics models via transfer learning
• Demonstrates Perch 2.0's effectiveness for few-shot marine mammal classification

Innovation

• Pretrained bioacoustics foundation model with 14,597 species
• Few-shot transfer learning for underwater audio tasks
• Linear probing with embeddings outperforms alternative models