Automated detection and annotation of toothed-whale whistles using transformer-based instance segmentation.

Area of Science:

• Bioacoustics
• Artificial Intelligence
• Marine Mammal Science

Background:

• Accurate dolphin whistle detection and annotation are vital for marine mammal communication research.
• Existing methods struggle with variable signals, overlapping calls, and noisy environments, limiting generalizability across species and whistle types.
• Manual feature engineering in traditional approaches is time-consuming and may not capture complex whistle characteristics.

Purpose of the Study:

• To develop an end-to-end AI system for accurate dolphin whistle detection and annotation.
• To improve the generalizability of whistle detection models to diverse species, locations, and acoustic conditions.
• To integrate a human-in-the-loop approach for iterative refinement of annotations and model performance.

Main Methods:

• Reformulated dolphin whistle detection as an instance-segmentation task using a transformer model.
• Developed an end-to-end system that predicts complete whistle contours directly from spectrograms.
• Implemented a human-in-the-loop training paradigm for iterative annotation refinement and model improvement.

Main Results:

• The end-to-end system demonstrated effective generalization on unseen species, locations, and hydrophones.
• Achieved 89.99% precision and 80.65% recall for all detected whistles.
• Attained 85.81% precision and 88.44% recall for whistles longer than 150 ms.

Conclusions:

• The proposed end-to-end transformer model offers a robust solution for dolphin whistle detection and annotation.
• The human-in-the-loop strategy enhances annotation quality and model performance, addressing limitations of manual labeling.
• This approach significantly advances the potential for automated analysis of marine mammal vocalizations.