I will talk about two related topics. First I will introduce the conditionally independent voice model, which expresses music as a collection of voices that evolve independently from one another when conditioned on a process that describes some shared evolving attribute such as harmony. I will show an application to pitch spelling from MIDI, though I believe the model may find use in a variety of musical applications. Attempts to train this model automatically using traditional Baum-Welch type methods were not particularly successful, due, perhaps, to the inappropriateness of the marginal likelihood criterion. We introduce a method for directly minimizing the error rate on a test set, using computational ideas from POMDPs.