I am having difficulty understanding certain concepts regarding the classification using HMM. There are numerous post here and in the internet about that, but they never get to detail or they all share some ambiguity.
My question is :
Lets say I train one HMM per class and now I would like to calculate the log likelihood for the test sequence to classify the new sequences. What data I will use here? is this the data the observation symbols that were not used in the training (test data) or would I first take this observation (test data) and pass it through the Viterbi algorithm and then find the log likelihood of that sequence (state sequence as opposed to observation sequence)?