Research On The System Of Query By Humming Based On E2LSH

Audio retrieval is becoming the new research hot spot since image retrieval has being developed in the field of content-based retrieval. As a branch of content-based audio retrieval, query by humming provides users with a new and convenient way of retrieval. When people forget the information such as the name of a song and so on, they can get the target rapidly in large-scale music database by the way of humming music clips.This thesis has a analysis of several technology of the modern music retrieval system based on humming, and we have a discussion and research on the representation of melody feature, the extraction algorithm of melody feature and the matching algorithm of melody. Today, most retrieval system of humming mostly ignores the speed. In this thesis, the technology of similarity retrieval based on Exact Euclidean distance locality sensitive hashing is applied to the retrieval of humming. On the basis, a fast system of query by humming is built. This thesis mainly does the following several aspects:1. The basis of the system of query by humming is conducted a analysis. Have a study of the method of the representation of melody feature and the relevance technology of speech signal processing. Research on the MIDI for creating the model of melody. And have a study of the extraction of melody’s feature, which includes pitch period forecast, smoothing and note segmentation. Compare several pitch period estimation method And YIN has a better performance. Then this thesis has a discussion on the problem about the pitch deviation and the problem about the duration scale in the process of melody matching.2. Research deeply and implement the technology of the melody matching. The performance of linear scaling(LS), dynamic time warping(DTW) and earth mover’s distance(EMD) are compared. For melody matching, Improvements such as dynamic threshold deviation constraint are come up and formed a faster algorithm which is called Improved DTW(IDTW). Combined with the advantages of EMD and IDTW, this thesis come up with a algorithm which is EMD-IDTW and whose performance of matching melody is more accurate and faster.3. Aimed at the problem of the modern music retrieval system whose melody matching speed is slow and which need exhaustive search, this thesis have a research on the Exact Euclidean distance locality sensitive hashing(E2LSH) with the theoretical analysis about its hash function. E2 LSH is applied to the system of query by humming, and the E2 LSH index structure based on the vector of pitch is created. With the index structure the retrieval speed of the system is increased. Finally, Combined with the EMD-IDTW, the Pitch-E2LSH+EMD-IDTW is come up and the system of query by humming based on E2 LSH is built, whose retrieval speed is improved greatly with guarantee high accuracy.