Hierarchical recurrent networks for learning musical structure

Layered neural networks employing feedback links have been proposed for certain sequential pattern tasks in automatic music composition. A hierarchical version of this type of network is studied. The use of such a hierarchical neural network for modeling coarse and fine temporal structure in music is investigated. This network is trained on two classical waltzes and then used to generate novel waltzes. The generated waltzes contained both novel phrases and phrases from the original scores. They exhibit an overall structure which has been difficult to learn using conventional methods. It is argued that it is the synaptic links of artificial neural networks which allow them to learn the relationship between coarse and fine temporal structure.<
>