Back to list

Automatic Detection of the Musical Structure within Pieces of Music

Authors Wankhammer, A.
Year 2011
Thesis Type Diploma thesis
Topic Audio Signal Processing
Keywords musical acoustics, Music Information Retrieval
Abstract Music Structure Discovery (MSD) for popular music is a well known task in Music Information Retrieval (MIR). In this thesis a new approach for finding the musical structure of a piece of music is proposed. The algorithm is based on the search for repeated vertical slices inside a modified bar level Self Distance Matrix (SDM) using a template matching algorithm. After an initial segmentation is found based on the analysis of the template matching results, a post processing step helps to further investigate the found musical structure by searching for repeated sub-sequences in the preliminary segmentation. The new idea of template matching instead of trying to find explicit blocks or off-diagonal lines inside the SDM is independent of any specific characteristics of the investigated ma- trix and can therefore be used on a wide range of different songs. The proposed algorithm has been evaluated on a well known dataset consisting of 13 full albums by The Beatles and the evaluation results have been compared to five state-of-the-art MSD algorithms. It could be shown that the segmentation performance of the proposed system on the chosen testing corpus is significantly better than the performance of any other system included in the evaluation. In a complementary experiment, the detected musical structure is used to investigate the influence of focusing the audio analysis to distinct representative song segments, when trying to find the genre label for a song. The results indicate that the additional knowledge can help to significantly improve the results of the classification system when focusing on representative sections instead of arbitrary parts, if very short excerpts (< 10 seconds) are used for classification. For longer excerpts, no difference for the classification of representative and arbitrary parts has been found.
Supervisors Sontacchi, A.