A new query-by-humming system based on the score level fusion of two classifiers

• Published in: International journal of communication systems Vol. 25; no. 6; pp. 717 - 733
• Main Authors: Nam, Gi Pyo, Park, Kang Ryoung, Park, Sung-Joo, Lee, Soek-Pil, Kim, Moo-Young
• Format: Journal Article
• Language: English
• Published: Chichester, UK John Wiley & Sons, Ltd 01.06.2012
• Subjects: dynamic time warping; linear scaling; quantized binary code; query-by-humming; score level fusion; spectro-temporal autocorrelation
• ISSN: 1074-5351; 1099-1131
• DOI: 10.1002/dac.1187

SUMMARY With the widespread use of multimedia devices, such as MP3 players, the necessity of a content‐based retrieval is increased, which can find the stored music even if a user does not know the title or singer of the music. Consequently, a query‐by‐humming (QBH) system is introduced, which provides functionality that a user can find a piece of music by humming. Although there have been many researches into QBH, there has been little done to combine more than two classifiers based on various fusion methods. Hence, we propose a new method of QBH based on the score level fusion of two classifiers. This research is novel in the following three ways as compared with previous works. First, the features of the humming data are extracted by using musical note estimation based on the spectro‐temporal autocorrelation (STA). We normalize the extracted features by using the mean‐shifting, median filtering, average filtering, and min–max scaling methods. Second, a pitch‐based dynamic time warping (DTW) method is used as the first classifier. We use the linear scaling (LS) method with the quantized binary (QB) code of the pitch data as the second classifier. Third, through the combination of these two classifiers based on the score level by the MIN rule, the performance of QBH is much enhanced. Experimental results with the 2006 MIREX QBSH and 2009 MIR‐QBSH corpus databases showed that the performance of the proposed fusion method was best compared with single classifier and other fusion methods. Copyright © 2010 John Wiley & Sons, Ltd. We extract the pitch from the humming file and delete all the pitch value of 0. Then, we normalize the features by using mean‐shifting, median filtering, average filtering, and min‐max scaling. These features are used for matching. Two scores are calculated by two classifiers such as the pitch based DTW and the QB code based LS method. Finally, the two calculated scores are combined by the MIN rule and the genuine MIDI file is searched for based on the ranking. Copyright © 2010 John Wiley & Sons, Ltd.