I'm trying to determine the state parameters for an hmm with a minimum of 9 states. I'm running chunks of my data through hmmtrain on MATLAB which I think uses the Baum/Welch aka forward/backward algorithm. This tends to be very slow. I have found several papers describing the segmental k-means algorithm for determining the hidden markov parameters. However, none of the papers make clear where in the algorithm the segmentation and k-means clustering occurs (see Rabiner 1990).
Any insight into this algorithm would be appreciated.
Thanks.