Lidia Mangu's Thesis

[james.waterhouse at elf.mcgill.ca]

Andreas,

Is there a paper that describes the algorithm used in SRILM?

Also, the reason I asked about Lidia Mangu's thesis and about the possibility of
a future reimplementation is because the current algorithm does not take into
account timing information. The research I am conducting is dealing with
recognition at the phone level. In creating sausages from the resulting
lattices it is uncertain how much timing information could effect the results.
Perhaps timing information may be more of an issue in phone recognition then in
word recognition.

Do you happen to know of any sausage creation software that takes into account
the timing information?

Thanks,
James


Quoting Andreas Stolcke <stolcke at speech.sri.com>:

>
> In message <1124912253.430ccc7d74da0 at www.webmail.mcgill.ca>you wrote:
> > Hello,
> >
> > This is a question for the developers of srilm... is there any plan to
> includ
> > e a
> > complete implimentation of Lidia Mangu's thesis in lattice-tools at any
> point
> >  in
> > the future?
> >
> > Thanks,
> > James
>
> James,
>
> the bad news is that there are no plans to reimplement Lidia's work in
> SRILM.  The good news is that it shouldn't be necessary.
> The SRILM lattice-tool program incorporates a new algorithm
> for confusion network construction from lattices.  The algorithm
> operates in time O(N*L), where N is the number of nodes in the lattice
> and L is the length of the confusion network.  The original algorithm
> can take time O(N^3) and is therefore much slower in most cases.
>
> This function is enabled by the lattice-tool -posterior-decode and
> -write-mesh
> options.
>
> I should note that several fast (linear-time) algorithms for
> CN construction have been published in recent years.
>
> --Andreas
>
>