Question about Lattice Tool

[Andreas Stolcke]

In message <3E8871D7.89295212 at itc.it>you wrote:
> Dear Dr. Andreas Stolcke and users of SIRLM
> 
> I have implemented our Lattice Tool based mainly on SIRLM tool kit (Lattice
> Tool) (our lattice definition has time information).
> My problem now is how to measure word error rate from the generated lattice
> file with the corresponding real utterance.  That means I need to compute the
> lower word error rate.
> Could you please tell some works about this problem before ?
> 
> In SIRLM Lattice Tool: there is a function namely:
> unsigned latticeWER(const VocabIndex *words,
>    unsigned &sub, unsigned &ins, unsigned &del)
>         { SubVocab ignoreWords(vocab);
>    return latticeWER(words, sub, ins, del, ignoreWords);
>  };
> seems to solve my problem. Unfortunately, I could not understand how this
> function work.
> I am looking forward to hearing from you.
> Thank you in advanced.
> 
> Vu Hai Quan.

Vu,

The lattice word error rate is computed by dynamic programming, very similar
to the standard word error computation on strings.  You keep a table 
that has the minimum error from the beginning of the string and lattice to
each pair of string position and lattice node.  For lack of time I cannot
describe the algorithm in more detail, but maybe someone on the list can
help with that.  It's really quite straightforward if you are familiar
with string alignment.  If you're not, then do a google search on
"string alignment algorithm" and you will find a dozen or so references
that should help you get going.

Incidentally, even if your lattices have time information it should be
irrelevant to the lattice error computation.  so you could convert
them to PSFG format and still use the SRILM tool.

--Andreas 

PS.  Please only send email related to the mailing list itself to
majordomo at speech.sri.com.  Once you've signed up, send all email to
srilm-user at speech.sri.com.