Polyphonic Voicing Optimization for Automatic Music Completion

In this paper, we present a new algorithm for automatic music completion. We have proposed automatic music completion as the class of music composition assistance problems of generating a complete piece of music given fragments of musical ideas input by a user. These fragments include partial melodies in multiple voices or parts of the underlying harmony progression. Therefore, it is a generalization of common problems such as melody harmonization or harmony constrained melody generation, but also includes problems with constraints in multiple domains, i.e. multiple voices and harmony. We present a new polyphonic voicing model for automatically completing four-part chorales. It is based on a hidden Markov model and what we call correction factors. These factors are trainable functions that efficiently capture the context of a voicing in order to account for a multitude of music theoretical rules without having to resort to a rule-based system. We observed improvement over our old model with regards to metrics that are derived from music theory for polyphonic voicing, and also invite the reader to try our algorithm themselves at http://160.16.202.131/music_completion_apsipa.