The Use of Large Corpora to Train a New Type of Key-Finding Algorithm

Abstract

Computational models of key estimation have struggled to emulate the accuracy levels of human listeners, especially with pieces in the minor mode. The current study proposes a new key-finding algorithm, which utilizes Euclidean distance, rather than correlation, and is trained on the statistical properties of a large musical sample. A model was trained on a dataset of 490 pieces encoded into the Humdrum “kern” format, in which the key was known. This model was tested on a reserve dataset of 492 pieces, and was found to have a significantly higher overall accuracy than previous models. In addition, we determined separate accuracy ratings for major mode and minor mode works for the existing key-finding models and report that most existing models provide greater accuracy for major mode rather than minor mode works. The proposed key-finding algorithm performs more accurately on minor mode works than all of the other models tested, although it does not perform significantly better than the models created by Aarden (2003), Bellman (2005), or Sapp (2011). Finally, an algorithm that combines the Aarden-Essen model (2003) and the proposed algorithm is suggested, and results in significantly more accurate key assessments than all of the other extant models.