Skip to Main Content
In this paper, we propose a novel approach for music identification with KD-tree and melody-line. In our method the process has three stages. Firstly, we use the features extracted from training data set to built a KD-tree. Secondly, features extracted from the music in the database, are quantified through the KD-tree into words. Then the words are stored. Meanwhile, the melody-line is also extracted from the music and also stored as a string. Thirdly, when the user gives a fragment song, features are extracted and then quantified the same way in the second stage, so is melody-line. We score the archive according to TF IDF scheme and get the best matches. String macthing of melody line is applied to re-arrange the orders of the best matches. Our contribution also includes a new kind of feature, MFCC Peaks, to acquire an efficient and accurate retrieval. The results of our experiments demonstrate that the accuracy of top5 is 98.54% while the top5 is 99.52%. We also compare our approach with Shazam algorithm and get higher accuracy among all six types of music.