Content-Based Music Similarity Computation With Relevant Component Analysis

Content-based music similarity is becoming important because of the millions of songs with online distribution. However, current methods have to treat the same attention on unrelated information or different informative information. In this paper, a new method is proposed to compute the music similarity with relevant component analysis. Considering the different weights for different parts, this method pays more attention on those parts with informative features by ranking the Mel-frequency cepstral coefficients frames on energy, and masters the song features more precisely compared with other methods. Experimental results on public dataset MusiClef show that this method works faster on music similarity computation task without sacrificing the accuracy of the similarity measurement.