SmartDJ, An Interactive Music Player for Music Discovery by Similarity Comparison

In this digital music era, sorting and discovery of songs is getting harder and more time consuming than before, due to the large pool of songs out there. Many music recommendation system and other similar applications in the market make use of collaborative filtering and social recommendation to suggest music to listeners. However, the problem arises when there is not enough information collected for the song, which happens mostly to new and less popular music. Other issues include missing or inaccurate metadata, the need for Internet connection, etc. We present research on acoustic features to automatically classify songs according to user-friendly and highlevel concepts that indicate social contexts for music listening, and a prototype application called "SmartDJ". We aim to provide novel ways that the user can browse her/his music collection, with a player that enhances interaction via a visual feedback, personalised DJ trajectories, smooth mix transitions and so forth. SmartDJ sorts the songs based on similarity by extracting low level features, then reducing feature space dimensionality with principle component analysis (PCA) and multidimensional scaling (MDS) methods, and plotting songs in a GUI for manual or automatic browsing, where song similarity is given by Euclidian distance in a lower-dimension song space. Users are able to visualise their music library and select songs based on their similarity, or allow the system to perform automation, by selecting a list of songs based on the selection of the seed song. Users can maneuver with the high-level descriptor on the interactive interface to attain the different song space desired.