Computerized Music Feature Analysis: Using MIDI Representation to Comprehensively Compare Musical Eras

Abstract

Music is a defining piece of our culture that has existed for centuries. The system of Western music that guides many styles we know today can be decomposed into fundamental units that can be systematically analyzed using Musical Instrument Digital Interface (MIDI) representation. Musicologists commonly agree that Western music has evolved over time, and hence it is pertinent to understand the underlying characteristics of musical eras to form a contextual basis for various compositional styles at different points in time. I utilize two corpora – the larger of which spans about 7,200 pieces and 1,700 composers – in order to examine three main musical features: pitch, harmony, and melody. Utilizing the music21 package, I first conduct pitch analysis to examine how tonality of composers’ works have changed over time. I then analyze harmony by calculating trichords and tetrachords and grouping them by tonality. Next, I use ABC notation to encode MIDI data and analyze several high-level melodic features. I conclude by compiling all of these features to train multi-class classifiers to predict time period of individual pieces, using accuracy, precision and other metrics to determine which musical eras are more or less distinct. I find a clear progressive shift from a predominant use of natural notes in the Baroque era to an even distribution of naturals and accidentals in the Romantic and Modern eras. Through examining triad types, I observe a predominant use of major triads over minor triads, and an increase in minor chords through the Romantic and Modern periods. Through melody analysis, I find that composers are utilizing increasingly large intervals through the four eras. After training five multi-class classifiers, I find that a Linear Support Vector Classifier achieves the highest accuracy of 60.1%. Examining confusion matrices for the two best models demonstrates that the classifiers are most likely to incorrectly predict a period that is temporally adjacent to the correct period, which implies a correlation between similar musical features and temporal proximity.