ChordNet: Learning and Producing Voice Leading with Neural Networks and Dynamic Programming

Abstract

We introduce a novel method for learning voice leading using neural networks. Unlike earlier approaches for learning chord sequences where chords are predicted from a local context, this method uses a multi-layer neural network for learning chord assessment from music examples. The network employs a special constraint-based topology for transforming the relative comparison of chord pairs into an absolute assessment function. Using this chord assessment function, globally optimized chord sequences can be found in linear time using the dynamic programming technique. The ChordNet implementation of the approach learns stylistic aspects of voice leading directly from a set of training examples. In combination with the HARMONET system, it presents a powerful framework for solving practical harmonization tasks.