Data-Driven Approach to Recognizing Influence Between Musicians

Abstract

All musicians have been influenced by previous generations of artists. In the past, composers would often pay homage to those that came before them by borrowing musical phrases or concepts. During the era of recorded music, the influence of older artists on younger musicians has continued at an even quicker pace. For example, James Brown, who was popular in the 1950's and 60's influenced Michael Jackson (popular in the 70's and 80's). Michael Jackson went on to influence Chris Brown (popular in the 2000's and 2010's). By examining the flow of influence, we can obtain an interesting look into how music has evolved and adapted throughout history. However, musical influence is a complex and nebulous concept. Such influence involves many components with varying degrees of importance, making it a very difficult topic to explore methodically. This stems from the many different types of relationships that could potentially be characterized as being influential to a musician. Musicologists have developed methods for assessing influence between composers of classical music, but these approaches tend not to be well suited for popular music, where the audio recording, rather than the musical score, is considered the canonical representation of the work. Therefore, I have attempted to explore musical influence, specifically among recording artists, by implementing a purely data-driven approach, with the hope of gaining a better understanding of the interplay between the various contributing factors. In this thesis, I have developed a preliminary system for modeling musical influences using a dataset of labels from the human editors of the AllMusic Guide. I also describe the development of a system that attempts to improve on that preliminary model by learning features for recognizing the influencing artists of a piece of music (from a selection of highly influential musicians). I also deployed a listening test to assess the performance of human listeners on the same task. The overall performance of the developed computational system is comparable to that of human listeners, supporting the data-driven approach presented in this work.%%%%Ph.D., Electrical Engineering – Drexel University, 2016