Music retrieval by singing and humming using information fusion

Abstract

We present that combinatorial fusion analysis (CFA) can improve results in a music information retrieval (MIR) task, specifically querying a database of recorded music by singing, humming, or whistling. Our experiment considers 10 scoring systems, 55 queries, and a database of 310 original artists' recordings. Through the use of spectral subtraction, we exploit the recording industry's tradition of placing the lead vocal and other prominent melodic features in the center of a stereo mix. We employ the rank/score function previously defined in other studies of CFA to analyze the behavior of scoring systems, and we use the rank/score variation to quantify the diversity of any two scoring systems. We then observe that successful 2-combinations, i.e. cases where the performance of a combination meets or exceeds the performance of its constituent scoring systems, tend to occur when each system performs relatively well and the systems are diverse.