Re: [Kaldi-developers] How to redo recognition on new lattice

[Arnab G. <ar...@gm...>]

Anna- if I understand correctly, you have made some modification to
the HCLG.fst and have separate decoding graphs for each utterance. In
that case you could simply call gmm-latgen-faster for each utterance:
you may have to use latbin/lattice-to-fst to convert the lattice to a
word graph and use a single line of the feats.scp as input. The other
option will be to write a specialized form of gmm-latgen-faster that
takes separate decoding graphs for each utterances (as an archive or
scp). You can see an example of such an approach in
bin/compile-train-graphs-fsts (compare to bin/compile-train-graphs).
-Arnab


On Sun, Aug 18, 2013 at 1:31 PM, Anna Schmidt
<an...@co...> wrote:
> Dear Kaldi developers,
>
> I am trying to combine Kaldi recognition results (in form of the result
> lattice that is output when decoding with gmm-latgen-faster) with
> situation dependent world knowledge which can also be represented as a
> lattice. For this I implemented a 'merging' of two lattices to produce a
> combined structure that is a lot smaller than HCLG, but should still
> contain the correct recognition hypothesis for the current sound file.
> We would like to redo the recognition on this structures to reduce WER
> further.
>
> So, my question is:
> Using Kaldi, how would it be possible to rerun the recognition on this
> new, combined lattice? There exists one unique combined lattice for each
> wav file in our test corpus since the situational knowledge is different
> for every file.
>
> Thank you very much for your help and all the best,
>
> Anna Schmidt
>
> ------------------------------------------------------------------------------
> Introducing Performance Central, a new site from SourceForge and
> AppDynamics. Performance Central is your source for news, insights,
> analysis and resources for efficient Application Performance Management.
> Visit us today!
> http://pubads.g.doubleclick.net/gampad/clk?id=48897511&iu=/4140/ostg.clktrk
> _______________________________________________
> Kaldi-developers mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-developers