Zero-Shot Respiratory Sound Classification via LLM Alignment
September 2, 2026
This framework aligns self-supervised respiratory encoders with medical terminology using LLM-synthesized reports to create semantic anchors. The method achieves a 61.3% mean zero-shot AUC across nine tasks, outperforming CLAP and Qwen2-Audio.
HOW THIS AFFECTS YOU
●
researcherYou can use LLM-synthesized semantic anchors to improve zero-shot performance in specialized domains.
●
healthYou can leverage foundation models for respiratory diagnostics without requiring large task-specific labeled datasets.