Visualizing similarity among estimated melody sequences from musical audio

We have developed a music retrieval system that receives a humming query and finds similar audio intervals (segments) in a musical audio database. This system enables a user to retrieve a segment of a desired musical audio signal just by singing its melody. In this paper, we propose a method to summarize the music database through similarity analysis to thereby reduce the retrieval time. The distance of chroma vectors is used as a similarity measure. The key. technique for summarization includes, mainly, a statistical smoothing method and a method of discriminant analysis. Practical experiments were conducted using 115 musical audio selections in the RWC popular music database. We report the summarization ratio as about 45%.