Salomea Grodzicka
supervisor: Piotr Bilski
The majority of currently utilized music streaming applications provide different kinds of song recommendation systems, vast majority of which are based on other user’s preferences (collaborative systems). The problem arises when it comes to signal analysis, due to signal-based features variations within a song as well as individuality of user’s music taste.
The system should employ the most efficient decision algorithm. During the research, the effectiveness of neural network classifiers will be verified (e.g. Bayesian Classifier, Nearest Neighbours Classifier, Support Vector Machines, Decision Trees or their combinations with gradient boosting).
The idea is to find continuous segments of each song (such as verse, chorus and bridge) and to perform signal analysis and decision algorithms on those consistent fragments. Another viable solution would be to take into consideration rhythmic patterns, oppose to most common recommendation content-based systems which use only danceability, tempo or power of the beat features. The next step could be to compute analogical features in tonal scale by using chromagrams.
The thesis will be introduced on an openly accessible dataset called Free Music Archive. It provides full-length and high-quality audio as well as pre-computed features, along with track- and user-level metadata and tags. Some of those audio features are introduced by Echonest (now well-known Spotify).