Systematic approaches to music rehearsal segmentation

Practice is always the secret to the success of musicians. Many musicians record their rehearsal sessions and listen back to reflect on their practice. However, the rehearsal sessions are unstructured, messy and rather long compared to commercial recordings. As a result, musicians may not have the capacity to listen back to all of their practice recordings comprehensively. Nowadays, Music Information Retrieval (MIR) techniques have been developed to better manage and filter informative representations from commercial recordings. However, not much research has been performed focusing on rehearsal sessions. What differentiates rehearsal session recordings from 'regular' recordings, is both their length and the unpredictability of the content within them. Besides, rehearsal recordings will not have labeled ground truth on their content, and obtaining this would require a massive amount of manual labeling work; thus, it is unrealistic to achieve. In this thesis, we therefore propose a systematic development and evaluation framework to deal with these challenges. In detail, we will focus on the segmentation of full rehearsal session recordings into meaningful repeated fragments that could be used by musicians.To this end, we propose a framework which adopts an unsupervised segmentation strategy that can be robust to expected variability in 'meaningful' repeats. While well-defined ground truth is absent, by employing an evaluation strategy that synthesises the ground truth based on real recordings, we still get insight into the performance of our methods