DeepAcoustics: Enhanced software for cetacean whistle repertoire analysis and categorization

Cetacean whistle repertoires provide insights into species identification, population structure, social behavior, and potential impacts of environmental changes. Accurately categorizing cetacean whistles into discrete acoustic types remains a fundamental challenge in repertoire analysis. The evaluation of whistle repertoires using advanced analytical approaches is complex, computationally costly, and fails to capture the nuanced distinctions of simple whistle types. DeepAcoustics, software modified from DeepSqueak with enhanced features for underwater tonal whistle analysis, groups whistle contours by applying a k-means clustering algorithm on a user selection of weighted input parameters, such as frequency, slope, and concavity. Before clustering, DeepAcoustics performed semi-automated whistle contour extraction. An integrated editing tool improved accuracy of traced contours and mitigated effects of noise on the intensity-based extraction algorithm yielding higher resolution contours. Contour characterization improved with increased sampling density by customizing the number of points per contour. Different k-means clustering cut-off methods were tested including user-defined clusters with minimum membership size and diagnostic tools such as batch silhouette and t-SNE plots were implemented. A systematic set of synthetic whistles helped gauge repertoire size estimation, analyzing optimal parameter selection and settings for cluster calculations. Challenges, successes, and opportunities for optimization will be presented.