The development and research of content-based music information retrieval (MIR) applications in the last years have shown that the generation of descriptions enabling the identification and classification of pieces of musical audio is a challenge that can be coped with. Due to the huge masses of digital music available and the growth of the particular databases, there are investigations of how to automatically perform tasks concerning the management of audio data. In this thesis I will provide a general introduction of the music information retrieval techniques, especially the identification of audio material and the comparison of similarity-based approaches with content-based fingerprint technology. On the one hand, similarity retrieval systems try to model the human auditory system in various aspects and therewith the model of perceptual similarity. On the other hand there are fingerprints or signatures which try to exactly identify music without any assessment of similarity of sound titles. To figure out the differences and consequences of using these approaches I have performed several experiments that make clear how robust and adaptable an identification system must work. Rhythm Patterns, a similarity based feature extraction scheme and FDMF, a free fingerprint algorithm have been investigated by performing 24 test cases in order to compare the principle behind. This evaluation has also been done focusing on the greatest possible accuracy. It has come out that similarity features like Rhythm Patterns are able to identify audio titles promisingly as well (i.e. up to 89.53 %) in the introduced test scenarios. The proper choice of features enables that music tracks are identified at best when focusing on the highest similarity between the candidates both for varied excerpts and signal modifications.

Top