Authors: Daniel Rönnow & Theodor Twetman
The aim of the study was to find a combination of machine learning algorithms and musical parameters which could automatically classify a large amount of music tracks into correct genres with high accuracy.
To mimic a real musical situation we used the Million Song Dataset as it contains pre-analysed data on a wide variety of tracks. On the basis of previous studies and our evaluations of the available musical parameters a selection of four algorithms and four combinations of parameters were made. All these combinations of parameters were evaluated with each of the algorithms.
The best algorithm used with the two best combinations resulted in 49% and 51% accuracy respectively. Compared to some of the previous studies in this field our results are not outstanding, but we believe our results are more relevant in a real musical situation due to our choice of dataset, parameters and genres. When we evaluated the parameters we discovered that the they differentiated very little between the genres.
Even though our results show that our implementation is not good enough to use in a real application, it does not exclude the possibility of implementing an application for automatic classification of tracks into correct genres with high accuracy. The fact that the parameters do not differentiate much indicate that it might be a very extensive task to achieve the goal of high accuracy.