Research On Audio Similarity And Their Applications In Music Retrieval

A great number of music works emerged on the internet in the big data era. The larger number of music works lead to information overload, which makes the user difficult to get their favorite music and the underground music works become increasingly deviated from the public view. Hence, it is needed to maximize the association between different music and feedback the user with music works that are similar to their preference. Research on music similarity and music information retrieval bears practical meaning.In this thesis, we focus on music similarity measure and music similarity retrieval, design a similar music retrieval system to retrieve similar music fast and accurately based on the user’s requirement for music similarity retrieval. Specific work and innovations of this thesis are listed below:1. Research on music similarity measureWe first study the classical music similarity measure: G1 algorithm, and then propose two improvements in the audio feature extraction and the nearest neighbor distance calculation. 1) Since the simplex feature extraction of G1 measure does not depict music attribute comprehensively, we introduce a linear combined feature extraction method, which depicts the acoustic characteristics of music from physical, perceptual and semantic aspects. 2) We propose a global scaling algorithm(GS algorithm) to ease the Hubs problem of music similarity measure. GS, based on the principle of pairwise stability algorithm, re-scale proximity matrix to reduce the Hub problem and improving accuracy of nearest neighbor computation.2. Research on music similarity retrieval.1) We propose the representative music similarity retrieval algorithm- G1-FR and improves both space projection and index structure building respectively. 1) Due to the unwilling bad projection caused by non-metric divergence, we use RJSD to calculate the distance between Gaussian features instead of K-L divergence to improve the performance of index structure in retrieval scenario. 2) Due to the limitation of existing technology that can only construct index structure for vector features, we use the fast map algorithm to embed the non-vector features to metric space, introducing FMLSH index construction algorithm for non-vector features combined with E2 LSH technique, which solves the disadvantages of exhaustive matching of existing similarity measure. 3) We introduce ML-GS-FMLSH music similarity retrieval algorithm, which combines ML-GS music similarity measure and FMLSH indexing structure. ML-GS-FMLSH could reduce the calculation complexity from O(N) to O(k) without precision decrease, where Nk ??. We then designs and implements a similar music retrieval system based on the algorithm.