Abstract
The Hawaiian monk seal is a phocid Carnivore endemic to the Hawaiian Archipelago. This species was thought to be silent under water until the recent discovery that mature males produce at least six distinct calls year round, typically in bouts of two or more, with sound production linked to reproductive status. Here, we apply a progressive deep‐learning approach for acoustic monitoring of this endangered marine mammal, based on vocalizations obtained from two adult seals housed separately in controlled settings. We used 4989 manually labeled calls and corresponding spectrogram images to fine‐tune a pre‐trained YOLO network via transfer learning to detect and classify sounds from long‐term recordings of the same individuals. The algorithm demonstrated 87% weighted precision and efficiently recovered 48,324 calls from 16,262 h of data. This dataset was then queried to confirm temporal trends in vocal activity and predictable patterns in call sequencing. We transitioned this approach to field settings by combining verified calls from the laboratory with soundscape recordings at a range of signal‐to‐noise ratios. The re‐trained algorithm recovered 2049 seal vocalizations from 12 h of recordings from critical habitats with varying background noise levels, successfully capturing the temporal patterns in acoustic presence documented by manual annotations. This machine‐learning tool enables the informed use of acoustic data to detect free‐ranging Hawaiian monk seals, supporting conservation efforts for this endangered species.