Baum-Welch Algorithm


I was reading the book by Jurafsky and this is written by the author on HMM

Although in principle the forward-backward algorithm can do completely unsu- pervised learning of the A and B parameters, in practice the initial conditions are very important. For this reason the algorithm is often given extra information. For example, for HMM-based speech recognition, the HMM structure is often set by hand, and only the emission (B) and (non-zero) A transition probabilities are trained from a set of observation sequences O.

My question is, what exactly is the author trying to say? What is the extra information given? And in the example, I don’t see any extra information. What does it mean to set the HMM structure by hand?