A Machine Learning Approach to Voice Separation in Lute Tablature

In this paper, we propose a machine learning model for voice separation in lute tablature. Lute tablature is a practical notation that reveals only very limited information about polyphonic structure. This has complicated research into the large surviving corpus of lute music, notated exclusively in tablature. A solution may be found in automatic transcription, of which voice separation is a necessary step. During the last decade, several methods for separating voices in symbolic polyphonic music formats have been developed. However, all but two of these methods adopt a rule-based approach; moreover, none of them is designed for tablature. Our method differs on both these points. First, rather than using fixed rules, we use a model that learns from data: a neural network that predicts voice assignments for notes. Second, our method is specifically designed for tablature—tablature information is included in the features used as input for the models—but it can also be applied to other music corpora. We have experimented on a dataset containing tablature pieces of different polyphonic textures, and compare the results against those obtained from a baseline hidden Markov model (HMM) model. Additionally, we have performed a preliminary comparison of the neural network model with several existing methods for voice separation on a small dataset. We have found that the neural network model performs clearly better than the baseline model, and competitively with the existing methods.