That gives you indexing
beta[k, i, j] if that's what you want, or
beta[k] to pick out a matrix.
The main problem with HMMs is that those matrices have to be simplex constrained if they're the transition matrices, so you want to use something like:
simplex beta[K, 3];
Same indexing structure,
beta[k, i, j], but now you are guaranteed to have a simplex. If you're defining the matrix yourself using something like softmax over predictors, this would only give you error checking.