Audio Foundation Models Capture Phylogenetic Signal in Marine Mammal and Bird Vocalizations

This paper tests whether pretrained audio embeddings encode phylogenetic structure that was not part of their training objectives. Across 32 marine mammal species and 20 bird species, general-purpose models recovered evolutionary relationships from vocalizations. For cetaceans, CLAP and BEATs-bio reached Mantel correlations of 0.82, while AST reached 0.74; handcrafted 105-dimensional MFCC features showed no significant signal. The result remained after dimensionality matching and partial control for dominant frequency. On birds, AST and CLAP again outperformed the domain-focused BirdNET and BEATs-bio, suggesting that domain-specific pretraining alone does not guarantee stronger biological representations.