Music thumbnailing for radio podcasts: A listener evaluation

Abstract

When radio podcasts are produced from previously broadcast material, 30-second "thumbnails" of songs that featured in the original program are often included. Such thumbnails are made up of continuous or concatenated sections from a song and provide the audience with a summary of the music content. However, editing full-length songs down to representative thumbnails is a labor intensive process, particularly when concatenating multiple song sections. This presents an ideal application for automatic music editing tools and raises the question of how a piece of music is best summarized for this task. To gain insight into this problem we asked 120 listeners to rate the quality of thumbnails generated by eight methods (five automatic and three manual). When asked to judge overall editing quality (on a five point Likert scale) listeners gave higher ratings to methods where the edit points were quantized to bar positions, although we found no preference for structural content such as the chorus. Ratings for two automatic editing methods (one containing the chorus, one containing only the intro and outro) were not significantly different to their manual counterparts. This result suggests that the automatic editing methods applied here can be used to create production quality thumbnails.