Hidden Markov Models: dealing with sequences of different length

Hi,

I'm training an HMM with a database containing sequences with very different lengths.

The update rule for the emision and the transition probabilities normalizes each contribution (emision or transition) from each sequence according to the posterior probability of that sequence (Eqs. 109 and 110 in Rabiner's tutorial).

This can be a problem when there are sequences with far different length, because the ratio among different sequences probabilities cannot be represented (it's above realmax).

However, when I go through the hmmtrain.m code, I find that such normalization is not being applied (line 267).

My question is: Is there a bug in Mathworks' hmmtrain.m or am I missing anything up?

Thanks in advance, Jose