Contrastive learning for passive acoustic monitoring: A framework for sound source discovery and cross-site comparison in marine soundscapes

TL;DR

This paper introduces a self-supervised machine learning framework for analyzing underwater sounds, enabling discovery of acoustic patterns without needing labeled data.

Contribution

The novel contribution is a contrastive learning framework that organizes marine sounds into meaningful clusters without manual labels.

Findings

The framework reveals recurring acoustic patterns corresponding to biological and anthropogenic sounds across multiple sites.

It produces more stable and coherent sound groupings compared to conventional feature-based and supervised methods.

The approach identifies both site-shared and site-specific acoustic signatures in marine soundscapes.

Abstract

Passive acoustic monitoring (PAM) is a powerful tool for studying marine biodiversity, but large-scale analysis of underwater recordings is constrained by noise, overlapping signals, and limited labeled data. Here, we present a scalable, unsupervised contrastive learning framework for marine soundscapes. Using a large PAM dataset spanning multiple biogeographies, we show that the proposed approach organizes recordings into clusters with well-defined internal structure, as assessed using intrinsic clustering metrics and within-cluster similarity. The resulting clusters reveal recurring acoustic patterns that correspond to broad sound-source categories, including biological sounds such as fish calls and choruses, and anthropogenic sounds such as vessel noise, without explicitly enforcing these distinctions during training. Compared with established approaches, including cepstral features, variational autoencoders, and supervised pipelines, the proposed framework produces embeddings that support more compact and stable unsupervised clustering while preserving fine-scale acoustic variation beyond predefined species labels. By learning a shared representation across recordings from multiple sites and years, we examine the reproducibility of acoustic patterns across locations and identify both site-shared and site-specific sound signatures. Although the method is not designed to recover coarse species labels, it enables label-efficient analysis by reducing reliance on manual annotation and supporting exploratory characterization of complex marine soundscapes. Together, these results highlight multi-positive contrastive learning with a teacher network and acoustically informed augmentations as an effective strategy for scalable, discovery-driven analysis of passive acoustic monitoring data.

Underwater ecosystems are rich with sound, ranging from fish calls and choruses to the noise generated by human activity, yet large-scale analysis of passive acoustic recordings is hindered by background noise, overlapping signals, and sparse annotation. We present a self-supervised machine learning framework that learns directly from reef acoustic recordings without requiring manual labels. Using contrastive learning, the model organizes sound fragments based on acoustic similarity, enabling unsupervised structuring of thousands of hours of audio. Applied to recordings from multiple Caribbean spawning aggregation sites, the approach revealed recurring acoustic patterns consistent with known fish vocalization activity, as well as site-specific sound types and distinct anthropogenic noise signatures. Compared with conventional feature-based and supervised methods, the proposed framework produces more stable and coherent acoustic groupings while remaining entirely label-free. These results illustrate how self-supervised learning can support scalable, data-driven exploration of passive acoustic monitoring data for characterizing marine soundscapes.

Figures