Harmonising Melodies: Why Do We Add the Bass Line First?

We are taking an information theoretic approach to the question of the best way to harmonise melodies. Is it best to add the bass first, as has been traditionally the case? We describe software which uses statistical machine learning techniques to learn how to harmonise from a corpus of existing music. The software is able to perform the harmonisation task in various different ways. A performance comparison using the information theoretic measure cross-entropy is able to show that, indeed, the bass first approach appears to be best. We then use this overall strategy to investigate the performance of specialist models for the prediction of different musical attributes (such as pitch and note length) compared with single models which predict all attributes. We find that the use of specialist models affords a definite performance advantage. Final comparisons with a simpler model show that each has its pros and cons. Some harmonisations are presented which have been generated by some of the better performing models.