TensorFlow Hidden Markov Model with more complex structure

Question

Using the great TensorFlow Hidden Markov Model library, it is straightforward to model the following Dynamic Bayesian Network:

where H_i is the probabil