But, we don't get observation sequences labeled with states. Hmm.
How about this? We just start with some guess for the HMM. Then we can label the sequence ourselves, compute the counts, and get a HMM.
Just twisted enough that it might work!
