Automatisierte Extraktion rhythmischer Merkmale zur Anwendung in Music-Information Retrieval-Systemen

This thesis describes the automated extraction of features for the description of the rhythmic content of musical audio signals. These features are selected with respect to their applicability in music information retrieval (MIR) systems. While research on automatic extraction of rhythmic features, for example tempo and time signature has been in progress for some time, current algorithms still seem to be a long way from matching human recognition performance. Among the reasons of the difference between the performances of a machine listening system and a trained listener are the use of information on different levels of abstraction and musical knowledge in human cognition. The approach described here is influenced by these two principles of cognition. In order to identify appropriate features and relevant aspects of human processing of audio signals the necessary knowledge of musicology, psychoacoustics and cognition science are described. Subsequently, the description of the state-of-the-art comprises known methods for the extraction of rhythmic features from musical audio signals. The main part of the thesis contains a collection of machine-listening methods evaluating information on different levels of abstraction. A compact representation of metrical structure of musical audio signals is proposed. The evaluation of low-level features enables the application of musical knowledge to a minimal degree only. On the other hand it becomes apparent, that the processing of high-level features is prone to errors due to the propagation of the errors in the extraction process of this information. This motivates the joint evaluation of lowand high-level information depending on their reliability. The extraction of rhythmic features from information of automated detected percussive instruments represents a technical progress compared to the state-of-the-art. The segmentation of the audio signals in characteristic and similar regions representing verse or chorus for example is introduced as a valuable pre-processing step. The achieved significant improvements of the recognition rate are proved with real-world test data. The performances of the developed methods are evaluated using a large corpus of test data and the applicability of the extracted features for the use in an exemplary MIRsystem is examined.