Audio foundation models capture phylogenetic signal without domain-specific training

A new study published on arXiv investigates whether large audio foundation models can capture phylogenetic signals from species vocalizations without explicit training for this purpose. The research found that general-purpose models like CLAP and BEATs-bio demonstrated significant ability to recover phylogenetic distance in marine mammal and bird species, outperforming hand-crafted features. Surprisingly, domain-specific models such as BirdNET did not show superior performance, suggesting that broad pretraining on audio data may be sufficient for encoding evolutionary information. AI

IMPACT Demonstrates that general audio models can encode evolutionary information, potentially impacting bioacoustics research and model development.

RANK_REASON Research paper published on arXiv detailing findings on audio foundation models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• BEATs-bio
• BirdNET
• CLAP
• Jetz et al. (2012)
• Mel Frequency Cepstral Coefficients
• Watkins Marine Mammal Sound Database

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →