Leopard seals (Hydrurga leptonyx) are widely distributed pack-ice seals in Antarctic and sub-Antarctic waters. As apex predators, these animals play a crucial role in the Southern Ocean food web. Data on population-level changes in their abundance and distribution may be a useful indicator of ecosystem-level changes in this unique and fragile environment. Over the past few decades, many studies have conclusively shown that passive acoustic monitoring (PAM) is an effective tool to monitor the abundance and distribution of vocally active marine mammals in remote and inaccessible areas for extended periods. However, handling and analyzing the vast amount of PAM data being collected remains challenging. Within the scope of this effort, we explored the use of a machine learning algorithm (convolutional neural network; CNN) to automatically detect the ‘low double trill’; one of the most common leopard seal vocalizations, in three years of continuous acoustic data recorded in the Bransfield Strait, Antarctica between 2005 and 2008. After optimizing the algorithm, we evaluated its detection performance on various temporal scales (weeks, days, hours) to assess if CNNs are useful for monitoring leopard seal populations at ecologically relevant scales in the Southern Ocean.