Learning and generating musical structures computationally is a challenging task that has taken the interest of a number of research projects since the start of modern computation. This research attempts to demonstrate the approach of using machine learning techniques, namely probabilistic models, to learn and generate large-scale musical structures. Musical works have several defining levels of representation, or musical parameters, such as metre, duration, phrase structure, cadential patterns, pitch. The concept of Hierarchical Input-Output Hidden Markov Models is introduced. These are used in combination with Hierarchical Hidden Markov Models to build composite networks of models that represent musical parameters. The aim is to learn both the local dependencies of the elements that make up the parameters, and the interdependencies between the different parameters. Structured probabilities are extracted from a musical data set, which are then used to generate new pieces of music, providing the models with only a minimum of supervision and expert knowledge. The musical material this study concentrates on is Bach chorales. The composite nature of the networks allows us to experiment with several combinations of models. In order to validate the approach, the generation of two part pieces is used as a preliminary test, later moving on to complete four part works. The musical results, created using a simple “random walk” method, are evaluated with a listening study and analysed using entropy values and music theoretical rules.