A Study On Query-by-Continous-Humming System Using Repeating Patterns

With the advance in multimedia and Internet technology, more and more multimedia information are available on the Internet. Many studies focus on how to get the multimedia information we want. The purpose of this paper is to study music information retrieval which is belong to the multimedia information retrieval, so, the user only need to hum part of a song and can get the song he or she wanted.The current studies on query by humming focus on humming in special way and approximate string matching. Little attentions are put on humming in continuous way, and even using the continuous input method, the matching is still based on the sequence of pitches which has not been processed first. To be more, because of number of data in the sequence of pitches is too much, the calculation requirement of melody matching module is very high. Therefore, how to improve the search speed of the system needs to be discussed.Based on such shortcomings in studies on query by humming, we did some work as followed: in the module of building a melody database, reduced the data that needed to be calculated by extracting the repeating patterns in the songs, and proposed an improved Repeating-Pattern-Tree(RP-Tree) algorithm, with much better performance; in the melody extraction module, firstly extracted the sequence of the pitches, then removed some wrong points and reduced the number of dimensions in sequence of pitches in some way after which a time series of pitches was generated; in the melody matching module, analyzed four familiar errors that occur to the humming and the requirement to the search algorithm, by which we designed a retrieval method of Dynamic Time Warping(DTW) algorithm based on N-Gram index; implemented a prototype system, and after doing the experiments, we got a top ten hit of 94%. The experiment result shows, the non-trivial repeating patterns of songs can not only reduce the data and memories in retrieval, but also reflect the music characteristics of songs. So, in the area of music information retrieval, the study of music structures and characteristics is very important and meaningful.