Researchers introduce MetaPerch, a new bioacoustic foundation model that leverages recording metadata such as location and time as auxiliary supervision signals alongside vocalizations. This approach allows the model to utilize species-metadata correlations, encouraging richer representations that generalize better to species distribution and acoustic domain shifts.
The study presents an extensive empirical analysis of the effects of 9 diverse metadata sources across 17 bioacoustic datasets.
By incorporating this additional information, MetaPerch achieves strong species identification performance across multiple challenging domains, addressing key deployment challenges in real-world passive acoustic monitoring.