Melody Harmonization With Interpolated Probabilistic Models

Most melody harmonization systems use the generative hidden Markov model (HMM), which model the relation between the hidden chords and the observed melody. Relations to other variables, such as the tonality or the metric structure, are handled by training multiple HMMs or are ignored. In this paper, we propose a discriminative means of combining multiple probabilistic models of various musical variables by means of model interpolation. We evaluate our models in terms of their cross-entropy and their performance in harmonization experiments. The proposed model offered higher chord root accuracy than the reference musicological rule-based harmonizer by up to 5% absolute.