Phylogenetic signal in marine mammal and bird vocalizations captured by audio foundation models: the limited benefit of domain-specific pretraining
Merged summary
TL;DR - General-purpose audio foundation models encode strong phylogenetic signals in marine mammal and bird vocalizations, substantially outperforming handcrafted features. Domain-specific bioacoustic pretraining provided limited additional benefit.
- CLAP and BEATs-bio achieved Mantel correlations of 0.82 among 26 cetaceans, while MFCC features showed no significant signal.
- The cetacean result persisted after dimensionality reduction and controlling for dominant frequency.
- AST and CLAP also recovered phylogenetic structure across 20 bird species, with correlations of 0.55 and 0.52.
- BirdNET and BEATs-bio did not outperform general-purpose models on birds, each reaching only about 0.32–0.36.
Sources (1)
Phylogenetic signal in marine mammal and bird vocalizations captured by audio foundation models: the limited benefit of domain-specific pretraining
TL;DR - General-purpose audio foundation models encode strong phylogenetic signals in marine mammal and bird vocalizations, substantially outperforming handcrafted features. Domain-specific bioacoustic pretraining provided limited additional benefit.
- CLAP and BEATs-bio achieved Mantel correlations of 0.82 among 26 cetaceans, while MFCC features showed no significant signal.
- The cetacean result persisted after dimensionality reduction and controlling for dominant frequency.
- AST and CLAP also recovered phylogenetic structure across 20 bird species, with correlations of 0.55 and 0.52.
- BirdNET and BEATs-bio did not outperform general-purpose models on birds, each reaching only about 0.32–0.36.