Musik-Genreklassifikation mit Hilfe von Hidden Markov Modellen

Music is part of our everyday life. With the help of car radios and mobile devices we listen to music from early in the morning until late in the evening, not only with on the way to work and school, but also during exercise, in restaurants and on vacation. Music information retrieval as a sub domain of information retrieval gains its popularity with the growth of the internet and the spread of music audio formats like MP3. In the last few years we have faced a lot of scientific work that has been published in this area and another kind of work which concentrates in development of efficient algorithms and databases for management of large music collections. A popular form of music organization is the classification of genres. The idea of automatic music classification has been born because of the need for this kind of organization and costs involved by manual classification.
In this work we inspect an often used method for music genre classification. Besides that, we try to give an overview of the music information retrieval, used techniques and also point to the sources of important information, data and existing implementations. The goal of this work is to study the music genre classification with help of Hidden Markov Models and the application of this to classify music provided at International Symposium for Music Information Retrieval (ISMIR) in the year 2004. The 30 seconds of music has been converted from MP3 compressed to wave audio format at CD-Quality. To extract the features we use the MIRToolbox implemented by the University of Jyväskyläda (Finland) - Department of Music as an installable toolbox for MATLAB.
Out of six extracted features, as a time-dependent characteristic trait of an audio signal, we build 63 different combinations, which are then trained and classified with help from Kevin Murphy's "Hidden Markov Toolbox" for MATLAB. Because we trained the models with three different training sets (33%, 50%, 100%), we could see how the size of the training set influences the classification with HMMs. The results here show, as expected, that a well chosen combination of features leads to better classification results then any feature individually and that Hidden Markov Models face the classification of music genres very well.
In this work we reached the classification accuracy of 75%.