Learning to Create Jazz Melodies Using a Product of Experts

We describe a neural network architecture designed to learn the musical structure of jazz melodies over chord progressions, then to create new melodies over arbitrary chord progressions from the resulting connectome (representation of neural network structure). Our architecture consists of two sub-networks, the interval expert and the chord expert, each being LSTM (long short-term memory) recurrent networks. These two sub-networks jointly learn to predict a probability distribution over future notes conditioned on past notes in the melody. We describe a training procedure for the network and an implementation as part of the opensource Impro-Visor (Improvisation Advisor) application, and demonstrate our method by providing improvised melodies based on a variety of training sets.