This article addresses the challenges faced by GLAM institutions in managing sound and audiovisual archives, characterised by exponential digital growth, funding constraints, and a shortage of specialised cataloguing expertise. In response to this “triple dilemma,” we advocate for developing AI-assisted metadata tools designed to augment—not replace—human expertise, thereby shifting focus from digitisation to knowledge organisation and recontextualisation. Through a case study on early 20th-century Chinese quyi(narrative singing) recordings, we demonstrate how a domain-specific AI model—integrated with a Retrieval-Augmented Generation(RAG) architecture and trained on classical texts and expert annotations—enables deeper semantic analysis and culturally sensitive description. Ultimately, we call for collaborative development of such AI systems among ethnomusicologists, archivists, and technologists. This human-in-the-loop approach aims to enhance the global accessibility and interpretability of sound archives while preserving the accuracy and richness of cultural contextualisation.

audio-visual colletctions ,  artificial intelligence(AI) ,  metadata ,  cultural interpretation ,  human-AI collaboration ,  sonic heritage ,  Retrieval-augmented generation(RAG) ,  ethnomusicology ,  knowledge graphs