New benchmark probes Audio LLMs' ability to hear beyond human limits

Researchers have introduced the Machine Hearing Benchmark (MHB) to explore the capabilities of Audio LLMs (ALLMs) in perceiving sounds beyond human audibility. The MHB includes 6,128 instances across 13 tasks and seven domains, featuring diverse input regimes with high sampling rates and long sequence lengths. Current ALLMs show limited zero-shot transfer to machine hearing, and supervised adaptation alone does not bridge this gap, indicating a mismatch in audio front ends. To address this, a lightweight encoder extension called MachEar was developed, which enhances a pretrained ALLM's machine-hearing abilities while maintaining its existing audio understanding. AI

IMPACT This research could expand the perceptual capabilities of AI models into new domains, enabling applications that monitor or analyze sounds outside the human hearing range.

RANK_REASON Research paper introducing a new benchmark and model extension for Audio LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• ALLMs
• arXiv
• Audio LLMs
• Hugging Face
• Liwei Liu
• Machine Hearing Benchmark

• paper
• other

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