Currently i can recognize isolated words using HMM through training an HMM model for each sign, and for a new word i take the sign for the model giving the highest likelihood.
When it comes to multiple words joint together, i don't know what to do? I know i should use Viterbi algorithm but i don't know how specially i don't know the start and end of each sign in the continuous stream also there're some unpredictable movements that come from the movement between the end of a sign and the start of the next sign
I'm using C# Accord.Net Library.
to be more specific my question is:
How can i train the HMM classifier to automatically decode the continuous sequence of "words" (and note that each one of the hmms represents a word) taking into consideration that in a live stream not every segment of a sequence is a word, some segments aren't words, if this is done through Viterbi algorithm , what's the supposed sequence of hidden states how can i get it, as based on my limited info the sequence length of hidden states is equal to the sequence length of observed states, and in this case each hidden state (word) will emit a sequence of observed points.
Thanks in advance