Soundslike - automatic content-based music annotation and recommendation for large databases

A manual indexing of large music libraries is both tedious and costly, that is why a lot of music datasets are incomplete or wrongly annotated. An automatic content-based annotation and recommendation system for music recordings is independent of originally available metadata. It allows for generating an objective metadata that can complement manual expert annotations. These metadata can be effectively used for navigation and search in large music databases of broadcasting stations, streaming services, or online music archives. Automatically determined similar music pieces can serve for user-centered playlist creation and recommendation. In this paper we propose a combined approach to automatic music annotation and similarity search based on musically relevant low-level and mid-level descriptors. First, we use machine learning to infer the high-level metadata categories like genre, emotion, and perceived tempo. These descriptors are then used for similarity search. The similarity criteria can be individually weighted and adapted specifically to specific user requirements and musical facets as rhythm or harmony. The proposed method on music annotation is evaluated on an expert-annotated dataset reaching average accuracies of 60% to 90%, depending on a metadata category. An evaluation for the music recommendation is conducted for different similarity criteria showing good results for rhythm and tempo similarity with precision of 0:51 and 0:71 respectively.