DIPLOMARBEIT Evaluation of New Audio Features and Their Utilization in Novel Music Retrieval Applications

With increased popularity and size of music archives – in both the private and professional domains – new ways for organizing, searching and accessing these collections are needed. Music Information Retrieval is a relatively young research domain which addresses the development of automated methods for computation of similarity within music, in order to enable similarity-based organization of large music archives. In music similarity many different aspects play a role, e.g. tempo, rhythm, melody, instrumentation, but potentially also the structure (chorus/verse), the lyrics and even the language of a song. Much research is done on the automatic extraction of those aspects in order to describe music semantically, without the need of manual annotation. Those feature extraction algorithms form the basis for a range of further tasks. Automatic organization of entire music archives into categories can be accomplished by the use of classification algorithms. However, often the definition of categories is a problem itself and thus methods have been created to cluster music collections solely by sound similarity. Clustering means that music which is very similar is grouped together and separated from music containing different characteristics. Visualizations have been devised to provide intuitive views of clustered music collections. This work contributes two new algorithms for automatic extraction of features from music and presents a number of improvements on an existing descriptor. It contains a study on the importance of considering psychoacoustics in feature computation. The new approaches are evaluated on a number of reference music collections as well as in international benchmarking events on music genre classification, artist recognition and similarity retrieval. Moreover, a set of novel applications for clustering music libraries on Music Maps is presented, allowing interaction with and retrieval of music both on personal computers and mobile devices. For demonstration of practicability Mozart’s complete works have been organized on a Music Map, the Map of Mozart, which has been created utilizing the previously evaluated audio descriptors.