Automatically detecting key modulations in J.S. Bach chorale recordings
Publication Type:Conference Paper
Source:SMC Conference 2011 (2011)
This paper describes experiments to automatically detect key and modulation in J.S. Bach chorale recordings. Transcribed audio is processed into vertical notegroups, and the groups are automatically assigned chord labels in accordance with Schonberg's definition of diatonic triads and sevenths for the 24 major and minor modes. For comparison, MIDI representations of the chorales are also processed. Hidden Markov Models (HMMs) are used to detect key and key change in the chord sequences, based upon two approaches to chord and key transition representations. Our initial hypothesis is that key and chord values which are systematically derived from pre-eminent music theory will produce the most accurate models of key and modulation. The music theory models are therefore tested against models embodying Krumhansl's data resulting from perceptual experiments about chords and harmonic relations. We conclude that the music theory models produce better results than the perceptual data but that all of the models produce good results. The use of transcribed audio produces encouraging results, with the key detection outputs ranging from 79% to 97% of the MIDI ground truth results.