Feature-Wise Indexing of Indian Light Music for Efficient Retrieval Using Brief Humming Queries

Query-by-Humming (QBH) is a well-established technique in music information retrieval, offering intuitive and user-friendly search capabilities. This study presents a robust QBH framework specifically designed for Indian Light Music (ILM), leveraging the Constant-Q Transform (CQT) to extract perceptually relevant spectral features. These representations are indexed using Scalable Nearest Neighbours (ScaNN), enabling efficient and scalable approximate similarity search across large audio corpora. The system effectively supports short-duration humming queries and retrieves corresponding full-length ILM tracks with high precision. Experimental evaluations demonstrate notable improvements in retrieval accuracy and computational speed compared to conventional methods. The proposed framework is resilient to pitch variations, imprecise queries, and the inherently diverse melodic structures characteristic of ILM.