Research On The Technology Of Cover Song Identification

Cover version music or cover song is a class of music with similar main melody. Due to differences in timbre, tone, speed and structure, the cover song identification becomes a very difficult research topic in the field of music information retrieval. Cover song identification tasks typically include two forms: 1) Enter a query song, it returns all determined cover versions from the database to realize music recommendation. 2) Given a pair of songs, it is determined whether they have a cover relationship to implement copyright identification.With the advent of the era of big data, recognition speed becomes a problem which should be solved, but the current cover song identification research mainly focuses on the recognition accuracy. In order to solve this problem, we introduce a fast recognition algorithm using time series shapelets. It can greatly improve the recognition speed and have a good accuracy at the same time. As we all know, music can be seen as a combination of voice and vocal accompaniment. Based on this fact, we have studied cover song identification algorithms based on Chroma feature and MFCC feature. In this thesis, the main research and innovative work are as follows:(1) Research on cover song identification based on Chroma feature. The algorithm of feature extraction and tone invariance are analyzed; comparing different dynamic time warping algorithms, it is concluded that in the field of cover song recognition, local optimization is superior to the global optimization, and we point out the algorithm have to consider the problem of structure invariance. We introduce the time series shapelets method to solve the problem and modify it according to the characteristic of music signal. 1) Considering the candidate subsequence exhaustive problem, we put forward the method of fixed window function which greatly reduces the number of candidate subsequence. 2) Due to the large number of cover song categories, the information gain is difficult to calculate. We use a criterion based on class distance. In this thesis the new algorithm is named music shapelets(MS), it reduces the complexity of computation obviously. As for the robustness of MS algorithm in some circumstances, we put forward the triple shapelets(Triple-MS) method. It has higher accuracy and the same time complexity with MS algorithm.(2) Research on cover song identification based on MFCC feature. The extraction process of MFCC feature was studied, in order to solve the problem of traditional MFCC feature recognition accuracy, this thesis carries on the improvement of the following: 1) The critical bandwidth changes with frequencies while Mel filter banks consider it insufficient, we redesign the bandwidth of filter banks based on the critical equivalent rectangular bandwidth. 2) For the sensitive degree of different frequencies, this thesis proposes weight factor. 3) Instead of hamming window, we adopt Blackman window to keep the harmony information. On the basis of the above, we get the EFCC feature, and the recognition accuracy was improved. Combining the advantages of Triple-MS and EFCC, we present the method of Triple-MS-EFCC which improves the recognition speed under the premise of accuracy.