BACH: An Automated Classifier of Musical Harmony

We propose BACH, a musical classifier that uses a Markov chain to analyze harmonic structure. Such a system may be used to sort or search for music by composer, period, or aesthetic pleasantness to a listener. It may also be employed in musical recommendation services such as Pandora or Last.fm or in systems such as Google Music or 8Notes. We first extract harmonically significant notes from a MIDI file, reducing chords to their roots to yield a monophonic harmonic line. We add a measure of cohesiveness to the harmony by assuming that each voice is normally distributed and moving notes with statistical significance to their appropriate voice. We then analyze the frequency of each melodic interval employed within the harmonic voice, weighting by duration according to Parncutt’s durational accent model, resulting in a 12x12 matrix representing the probability of a transition from the chromatic row degree to the column degree, relative to the key of the piece. Employing the Euclidean norm as a distance measure, we perform k-nearest neighbor classification. Classification results on Beethoven and Bach piano pieces were excellent, illustrating the efficacy of our approach. Future work will focus on extending this approach to the problem of automated musical composition.