Design and Realisation of an Efficient Content Based Music Playlist Generation System

This thesis is on the subject of content based music playlist generation systems. The primary aim is to develop algorithms for content based music playlist generation that are faster than the current state of technology while keeping the quality of the playlists at a level that is at least comparable with that of the current state of technology. Not only need the algorithms be fast, they shall also allow flexibility for the end user to be able to tune the algorithms to match his personal requirements. For evaluation of the algorithms, a framework for automatic content based music playlist generation is developed and presented. In order to be able to evaluate the quality of music playlist generation systems, criteria for quality judgment of playlists have to be known. To gain insight in these quality criteria, a questionnaire is developed. The responses on this questionnaire are analysed. It shows that the number of parameters that influence the perceived quality of a personal playlist is huge, and the individual variation of desired values is large. Because of the large variance in preferred values, it is impossible to find one single set of parameters that suits for all people. Using the results of the questionnaire, it is argued that playlist genre consistency is a suitable criterion for assessing playlist quality. Songs within a playlist should have approximately the same genre. The key to good music playlist generation systems therefore is a good music similarity measure, that allows finding ‘similar’ music. To speed up music playlist generation systems, the music similarity measures used by these systems should be fast. This thesis presents two steps towards faster music similarity measures. The first step is early in the process of determining music similarity. The properties of each song are statistically described by a Gaussian mixture model. Similarity between songs is determined by applying the earth mover’s distance on these statistical models of the songs to measure the similarity of the songs. Where the current state of technology uses a constant number of Gaussians for all songs, this thesis presents a method to estimate an optimal number of Gaussians for each individual song. This prevents the Gaussian mixture