Abstract:

Evolutionary studies have become a dominant thread in the analysis of large audio collections. Such corpora usually consist of musical pieces by various composers or bands and the studies usually focus on identifying general historical trends in harmonic content or music production techniques. In this paper we present a comparable study that examines the music of a single band whose publicly available live recordings span three decades. We first discuss the opportunities and challenges faced when working with single-artist and live-music datasets and introduce solutions for audio feature validation and outlier detection. We then investigate how individual songs vary over time and identify general performance trends using a new approach based on relative feature values, which improves accuracy for features with a large variance. Finally, we validate our findings by juxtaposing them with descriptions posted in online forums by experienced listeners of the band's large following.

Direct link to video