MASTERARBE IT Discriminant Analysis of Three Rhythmic Descriptors in Musical Genre Classi cation

The introduction of digital music representation considerably altered the ways of creating, accessing and using music. Until today an immense number of music archives has been made available so that the actual attitude of music consumption has changed fundamentally. Both the commercial domain as for instance represented by music producers or music distributors and the private domain play a major role in the increasing importance of digital music archives. Yet, the size of music archives which can often be enormous demands new requirements according to the internal organization of included musical pieces as well as the individual access and search of musical pieces. Consequently, this means that scalable methods must be provided to automatically establish organizations of music archives according to speci c musical semantics. The research eld of Music Information Retrieval (MIR) aims to develop such methods which make possible a grouping, i. e. clustering or classi cation, of music pieces according to speci cally de ned musical semantics. Basically, such a musical semantics refers to the measuring the similarity of the underlying musical content. The de nition of this content-based similarity is based on individual musical characteristics like for instance rhythm, melody, instrumentation or others. Musical genres represent a very popular and frequently used musical category to organize music collections. In comparison to other possible musical categories, musical genres provide an intuitive understanding for categorization and are frequently used by humans to organize music. For instance, music retailers or music libraries widely use genre categorization to provide an e ective organization of o ered music collections. Within the MIR community the assumption generally holds true that the understanding of genres is potentially based on the descriptive power of certain content-based characteristics of the included musical pieces. Consequently, speci c genres may be actually related to a certain rhythmic, melodic or other musical characteristics. Unfortunately, this assumption of music genre representation based on content-based semantics appears to be insu cient as not content-based characteristics like for instance the cultural origin of artists and the cultural context of lyrics also play a role in the de nition of musical genres. Based on that potentially descriptive power of genres this master thesis examines the discrimination of musical genres in terms of rhythmic characteristics. Since various rhythmic descriptors exist in MIR, the three descriptor Rhythm Pattern (RP), Statistical Spectrum Descriptor (SSD) and Rhythm Histogram (RH) have been used throughout this thesis only. Each of these three descriptors contains a large number of features to constitute the speci c rhythmic component of