Abstract
Maritime audio monitoring plays a crucial role in marine traffic management, environmental protection, and marine life conservation. While recent advances in underwater acoustic signal processing have leveraged deep learning architectures like CNNs and transformers, most models are trained on clean or lightly noisy data and struggle under realistic, acoustically cluttered conditions where anthropogenic and biological sources overlap. This work presents a progressive enhancement of the Listen, Think, and Understand (LTU) audio-language model to address these challenges in comprehensive maritime soundscape analysis. Our methodology follows a three-stage progressive fine-tuning approach. First, we establish baseline performance by fine-tuning LTU for basic ship classification. In the second stage, we advance beyond simple classification to temporal segmentation, enabling the model to identify not just what marine vessels are present, but precisely when they occur in the audio stream. The final stage extends the model's capabilities to simultaneous detection of anthropogenic (ship) and biological (marine mammal) sounds, directly addressing the challenge of acoustically cluttered maritime environments. Our experiments demonstrate that this progressive fine-tuning approach, combined with realistic mixed-source training data, enables robust performance in acoustically cluttered conditions that challenge traditional models. The final system provides capabilities for automated maritime surveillance and marine ecosystem monitoring in underwater environments.