There have been many attempts at developing generative music systems. Rule-based methods generate a limited range of consistent but repetitive music. Probabilistic models are more flexible, allowing the generation of diverse music but lacking the clear causal structures and hierarchies found in most music. We are developing a generative system that induces the qualitative characteristics of motion in music by hierarchically analyzing and modeling changes of momentum simultaneously along multiple parameters, and subsequently transforming our model to give rise to variety and surprise.