A Task-Aware Compression Framework for Efficient Transmission of Marine Mammal Embeddings in Bandwidth-Constrained Acoustic Networks

Abstract

Real-time passive acoustic monitoring is fundamentally limited by the constraints of underwater telemetry. Acoustic modems typically provide bandwidths of only tens to hundreds of bits per second, which are insufficient for transmitting the high-dimensional acoustic data generated by monitoring systems in real time. Modern audio foundation models achieve state-of-the-art classification accuracy but generate high-dimensional embeddings that are operationally impractical to transmit from edge deployments. This paper addresses the critical trade-off between accuracy and bandwidth, proposing a method to enable verifiable, high-performance monitoring within strict transmission budgets. We propose a task-aware compression pipeline evaluated on a curated 15-class subset of the Watkins Marine Mammal Sound Database. We benchmark three pre-trained audio embedding models, identifying a bioacoustic foundation model as the superior but overparameterized feature extractor. We employ a supervised bottleneck autoencoder to distill task-discriminative features with varying loss functions, followed by product quantization and optimized product quantization to minimize the bit-rate while preserving class separability. Our approach achieves extreme compression ratios of 256 to 768 times while retaining over 97 percent of the initial baseline classification accuracy. Experimental results demonstrate that classification on the compressed latent representation matches or surpasses full-embedding performance. Information-plane analysis attributes this to the removal of recording-condition artifacts that caused mild overfitting in the uncompressed baseline. These findings demonstrate that the generalization tax of large foundation models can be dramatically reduced for domain-specific applications. The proposed framework enables practical marine monitoring systems that combine the accuracy of foundation models with the efficiency required for underwater acoustic networks.