A Query-By-Humming Based On Melody Feature

Content based multimedia retrieval is now one of the most popular subject in multimedia retrieval. One of content based music retrieval is query-by-humming, which is also the most straightforward way.There are totally three modules in query-by-humming system, i.e. processing of users'queries, construction of music database, and the matching module. Most of researchers put their emphasis on the matching module, few researchers pay more attention to the other two. The main content of this paper is about these two modules. For users'query processing, the most important task is melody extraction from humming clips. In this paper, we presented two methods to solve this problem, i.e. melody extraction based on rules and melody extraction based on Semitone Frequency Coefficient. The main idea of the former is to analyze the location of pitch in spectral. Comparing to some other methods, such as Autocorrelation, our method is more efficient and effective. The main idea of the latter is to calculate the probability of each semitone to belong to the main melody, and then the optimal melody contour is obtained. The experiments show that the former method fits processing of users'queries, and the latter fits melody extraction from polyphonic songs. In constructions of music database, we did research on MIDI database and polyphonic database. For MIDI database, we proposed a method based on rules to identify the main channels of MIDI music. For polyphonic database, we use Semitone Frequency Coefficient and Viterbi Algorithm to track the melody contour. Then a two-stage indexing scheme is constructed, where the first is based on tune features of melody and the second is based on profile features of melody. In matching modules, we first filter the candidate pool via the two-stage indexes, which can reduce the number of candidates. Then Dynamic Time Warping method is used to find 10 most similar songs from music database.In the research of query-by-humming, the experimental results are given in the form of rank lists. The experiments show that our method can make the TOP-3 hit rate to reach 71%, TOP-5 to 88% and TOP-10 to 94%.