Re: [Kaldi-users] Memory requirement for FSTs

[Vassil P. <vas...@gm...>]

There are people on this list that are far more knowledgeable than me,
and perhaps someone can give a more complete/correct answer, but my
understanding is that it boils down to the search space
representation. In the traditional recognizers the search space is
expanded on-demand, whereas with the (vanilla) WFST approach the
entire search graph is statically built before the recognition begins.
As far as I know there are ways to construct the graph on the fly, but
I am not sure if something like this is currently implemented in
Kaldi(I think not). Depending on your needs/constraints you may want
to have a look at say PocketSphinx, although on the other hand Nuance
released a WFST based recognizer for Android some time ago(not sure
how well it works): https://github.com/android/platform_external_srec
.

Vassil

On Thu, Oct 17, 2013 at 10:26 AM, E <oth...@ao...> wrote:
>
> Hello,
>
> I have been playing with Kaldi online recognizers (great work!) and wanted
> to ask if the FST approach is useful if I'm running under memory
> constraints. If I use traditional ARPA language model + acoustic models;
> total size of models is < 100 Mb (for 20,000 vocab size). But the HCLG.fst
> takes a whooping 500 Mbs! Why is this so (Perhaps I should read the papers
> to find the answers, but in short why size of HCLG.fst >> sum of size of
> individual *.fsts)? Is there some redundancy involved?
>
> What might be alternatives if one want to further reduce the size of
> HCLG.fst?
>
> Thanks.
>
>
> ------------------------------------------------------------------------------
> October Webinars: Code for Performance
> Free Intel webinars can help you accelerate application performance.
> Explore tips for MPI, OpenMP, advanced profiling, and more. Get the most
> from
> the latest Intel processors and coprocessors. See abstracts and register >
> http://pubads.g.doubleclick.net/gampad/clk?id=60135031&iu=/4140/ostg.clktrk
> _______________________________________________
> Kaldi-users mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-users
>