Corpus-Derived Key Profiles Are Not Transpositionally Equivalent

A fundamental assumption of distributional key-finding methods is that the frequency distributions of pitch classes in all keys are transpositionally equivalent. We tested this assumption with three experiments. First, using data from the openings of 995 major-key pieces and 596 minor-key pieces in the Yale-Classical Archives Corpus, we found that scale-degree distributions differ significantly from one key to another, and further analysis revealed that pieces keys with signatures having relatively more accidentals exhibit significantly more chromaticism than keys with fewer accidentals. Second, we examined whether these data might be accounted for by different keys’ varying modulation tendencies, and found this to be the case: keys with more accidentals modulate more frequently to more distant keys. Finally, we attempted to exclude modulatory passages from our data using a key profile analysis to identify key and mode within our dataset; however, the results of Experiment 1 still held. In sum, even when using a method that assumes transpositional equivalence, we found a difference between key profiles of different keys.