Similarity Measurement Method between Two Songs by Using the Conditional Euclidean Distance

Since numerous songs have recently been released increasingly, the genre of the song clustering is reasonably more important in terms of the audience’s choice. Also arguments for plagiarism are continuously being raised. For this reason, similarity measurement between two songs is important. In previous works, although similarity measurement has been actively researched in the field of query by humming, they only focused on quite partial matching for input humming. To solve this problem, we proposed a novel similarity measurement method between two songs. The proposed method has several advantages compared with previous works. Firstly, it is possible to measure overall similarity between two different songs. Secondly, overall region of a song is represented as 1-dimensional signal which can be obtained by run-length representation of 2-dimensional note information ((pitch, duration)). Thirdly, by sequentially adopting median filter, average filter, Z-score normalization into the 1-dimensional signal, we obtain the overall flow without noise feature such as the eccentric note of the song. Lastly, a new distance metric namely the conditional Euclidean distance is used by combining two distance concepts such as the Euclidean distance and the Hamming distance. To perform the feasibility test, several famous songs by the Beatles and the MIREX`08 dataset were used for our experiment. Also, by applying our method into a comparison between two songs with a plagiarism issue, we confirmed that very high similarity score between the two songs was measured. Key-Words: Music similarity measurement, Music clustering, Music plagiarism