Using Thumbnail Length Bounds To Improve Audio Thumbnailing For Beatles Songs

Abstract

Optimising the parameters of the audio thumbnailing procedure can improve the final results. Previously, experiments with the thumbnail length parameter have shown strong potential to enhance thumbnail boundaries detection for Beatles songs. However, usage of the thumbnail length parameter has been limited to only changing the thumbnail length lower bound. The purpose is to use the thumbnail length upper bound in combination with the lower bound to improve thumbnail boundaries' detection for Beatles songs. I experiment with the thumbnail length upper bound while fixing the lower bound, then analyse the F-measure results based on segment boundaries. I use a thumbnail procedure with a repetition-based fitness measure as the foundation. The results demonstrate that the thumbnail length upper bound can increase an estimated thumbnail boundaries' accuracy for Beatles songs. I select a pair of lower and upper bounds that slightly improves the F-measure based on segment boundaries, unlike using only the lower bound. In conclusion, this study optimises the thumbnail length bounds to improve the audio thumbnailing procedure with a repetition-based fitness measure for Beatles songs. It is demonstrated that the upper bound can improve the F-measure if chosen correctly. Unexpectedly, the upper bound can be omitted without losing much in the accuracy of thumbnail boundaries' detection. Additionally, I indicate further directions to optimise thumbnail length bounds for popular music and its genres (like pop, rock). Also, I describe other supplemental tasks for future work.