Re: [Cmusphinx-devel] Sphinx3j progress

[Bhiksha R. <bh...@me...>]

Philip here has it nicely stated :-)
You only need to ensure that you dont create duplicates of HMM states.
The hash will take care of the problem.

There wont be a problem during decoding, of course, since tokens have
time stamps and a token with a given time stamp can visit a state only once...

-Bhiksha

Philip Kwok wrote:

> Yes. HMM states that have a transition to itself can cause infinite loops in the
> way you described. The other case is that within the HMM, two HMM states can
> transition to the same third HMM state. Following the algorithm you described
> can cause the third HMM state to be duplicated. So the SentenceHMM builder will
> somehow have to remember that an HMM state has already been added. If it has
> already been added, it should simply just add an arc to it (instead of creating
> a new HMM state).
>
> The same thing for GrammarNodes. The traverser that dumps the current
> SimpleGrammar uses a hash to remember which SimpleGrammarNode is already
> visited. This way, these node won't be visited again, avoiding the infinite
> loop.
>
> phil
>
>         Given that I have a Unit for a pronunciation of a word, I need to add to
>         the sentence hmm the set of HMM states for the word.  To do that I get
>         the appropriate HMM from the Acoustic Model for the unit. I then get the
>         set of initial HMMStates (the states that have non-zero transition
>         probabilities from the initial state).  I then add a SentenceHMMState
>         for each of these initial HMMstates. Next, for each sentence hmm state I
>         get its successes by looking for non-zero transition probabilities and
>         add a sentence hmm state for each of these hmms. This process continues
>         until I reach the end of the unit.  The problem is that one of the
>         transitions out of an hmmstate is to itself. If I blindly follow the
>         algorithm and create a new sentencehmm state for all successors, I will
>         end up looping forever adding state after state for the hmm.  Perhaps
>         all that is needed to be done is to recognize the self transition and
>         have a similar self-transition in the sentence hmm.  This will mean,
>         though that the sentencehmm itself will be cyclical and and code that
>         traverses this tree will need to be wary not to fall into loops. The
>         search itself should be ok since we can rely on the pruner.
>
>         Thoughts on this?
>
>
> _______________________________________________________________
>
> Have big pipes? SourceForge.net is looking for download mirrors. We supply
> the hardware. You get the recognition. Email Us: ban...@so...
> _______________________________________________
> Cmusphinx-devel mailing list
> Cmu...@li...
> https://lists.sourceforge.net/lists/listinfo/cmusphinx-devel