Acoustic metrics (AMs) aggregate the acoustic information of a complex signal into a unique number, assisting our interpretation of acoustic environments and providing a rapid and intuitive solution to analyze large passive acoustic datasets. Manual identification and characterization of intraspecific call trait variation has been largely used in a variety of sonic taxa. However, it is time consuming, relatively subjective, and measurements can suffer from low replicability. This study assesses the potential of using a combination of standardized and automatically computed AMs to train a supervised classification model, as an alternative to discrimination protocols and manual measurements to categorize humpback whale (Megaptera novaeangliae) song units from the Southern Ocean. Our random forest model successfully discriminated between the 12 humpback whale unit types (UT), achieving an average classification accuracy of 84%. UTs were further described and discussed in the context of the hierarchical structure of humpback whale song in the Southern Ocean. We show that accurate discriminant models based on relevant AM combinations provide an interesting automated solution to use for simple, rapid, and highly reproducible identification and comparison of vocalization types in humpback whale populations, with the potential to be applied to both aquatic and terrestrial contexts, on other vocal species, and over different acoustic scales.