If you're training the HMM on one long string then it only has one example of a transition from the start state, thus your initial transition probability is rather meaningless. To get a meaningful estimate of $\pi$, you must break the sequence into segments and precede each segment with a distinguished 'start state'.

If you're training the HMM on one long string then it only has one example of a transition from the start state, thus your initial transition probability is rather meaningless. To get a meaningful estimate of $\pi$, you must break the sequence into natural segments and precede each segment with a distinguished 'start state', then train the model.