I'm designing a multi class classifier (for 4 classes) using Discrete HMMs with States N and Symbols M for each of the HMM.
However, I found that recognition performance(i.e highest log likelihood) of a specific class depends upon the N and M with which it was trained.
Most of the literature I found on the web suggests use of same of number of states, symbols across all the HMMs.
I was wondering if it need not be so? i.e. HMM1 can have 4 states, 18 symbols and HMM2 can have 5 states, 20 symbols............