Learning to analyse tonal music

Abstract

This work is an effort towards the development of a system for the automation of traditional tonal analysis of polyphonic scores in symbolic format. The system detects chords with their tonal functions, and key changes. All the possible tonal and key analyses are represented as a weighted directed acyclic graph. The best analysis is the path that maximizes, through a dynamic programming algorithm, the sum of weights in the graph. The selection of the weights according to the importance of each possible harmonic progression is a key issue. A genetic algorithm is proposed to learn them from a training corpus of a given music style. The proof of concept of this approach has been tested on Bach chorales.