MelTS Melody Translation System

MelTS is an automatic harmonization system that creates multi-part arrangements in the style of the data on which it was trained. The system approaches the problem of harmonization from a machine tranlsation perspective, modeling the melody of a song as the source language and each harmony as a target language. The approach stands in contrast to previous approaches to the harmonization problem, which have primarily taken two forms: satisfying rules provided by music theory or predicting the chord under the melody. A major benefit of the MelTS approach is that the style of the harmony voices can be learned directly from the training data, just as vocabulary and structure can be learned from a parallel corpus of natural languages. In particular, generating harmony lines probabilistically and individually, as opposed to by rule satisfaction or as a consequence of a predicted chord, allows for the production of more natural arrangements without the need for simplifying assumptions about the structure of the music to be produced.