Effective and Efficient Melody-Matching Method in a Large-Scale Music Database

Abstract

Rapid progress in hardware and software technologies has made it possible to manage and access large volumes of music data. The effectiveness (high-accuracy) and efficiency (high-speed) of retrieval techniques for music information data are key issues in the area of large-scale MIR systems. Retrieval performances by conventional methods, however, have not been good enough in large-scale music information retrieval systems. DP matching generally has high retrieval accuracy. However, its retrieval time depends on the size of the database. Thus, to improve the effectiveness and efficiency of music information retrieval in large-scale music databases, two techniques propose: fuzzy-quantization based on melody representation and error-tolerant linear matching. The fuzzy-quantization method uses membership functions based on a probability distribution. The error-tolerant linear matching method is a linear-matching method that considers the possibility of the insertion or omission of notes. Using these techniques, higher recall rates were achieved compared to conventional methods. Retrieval time in pre-selection was about 0.5 seconds per query in large-scale DB, including database loading.