The classification error probability of your model is inversely proportionally given by its likelihood, which is internally used by the Baum-Welch algorithm in order to improve your model at each iteration.
Just be aware that any decent BW implementation uses -log-likelihood in order to avoid underflow.
Now, if you want to go a step further, read this paper: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=5983107