BEANS-Next and ROOTS: Broadening Audio-Language Capabilities for Bioacoustics
Researchers introduced BEANS-Next, a benchmark testing broad bioacoustic audio-language abilities across four tiers: acoustic perception, biological recognition, scene understanding, and in-context learning. They also created ROOTS, a large training dataset of 44 million audio-language pairs. Training on ROOTS significantly improves model performance across all these tiers compared to existing models.