How to smooth the observation probability distributions of trained HMMs

[ganji]

Dear nickgillian,

I want to use DiscreteHMM to classify the handwritten words around 2,000 categories, however, the recognition rate is really poor (around just 46%). Fortunately, this paper (http://www.sciencedirect.com/science/article/pii/S0031320300000510) shows that the performance can be
improved if smoothing the observation probability distributions of trained HMMs. So, I want to know is it possible to implement this smoothing procedure in GRT? Or can I imporve the recognition results with other alternative?

I am looking forward to your early reply.
Thanks!