Evaluating impact of audio feature configuration on perceived quality of music rearrangement systems

Sammanfattning

A common task when creating videos is finding the right music to accompany your video. However, once an appropriate song is found it must usually be edited in some way or form to fit with the video. Automatic music rearrangement is the task of producing a new version of an existing song that has an either longer or shorter duration and is exclusively made up of unedited parts from the original audio. Proposals in the literature rely on audio feature extraction which are used to compare the individual beats of audio in the song (sometimes referred to as bixels). Given a desired target duration, the most-similar beats are chosen in a sequence which make up a new version as close to the target duration as possible. The features are joined in a linear combination using weights, most often equal weights are used. We wish to investigate the impact of different weights on the perceived quality of a music rearrangement system. We survey the available tools for music rearrangement and create a simple music rearrangement tool, which we also make available. We conducted a listening study using different weight combinations to gather subjective feedback on perceived quality. The results were in line with our expectation that all examined weight combinations produce good versions. There was no noticeable preference for any of the feature combinations, further indicating that all features were performant in the system we had built and that suitability varies from song to song. We also discuss how to best evaluate this subjective task and propose future direction for research on this topic.