Re: [Kaldi-users] How to improve the performance by adjusting Kaldi decoding parameters

[Mailing l. u. f. U. C. a. U. <kal...@li...>]

>        I am going to compare the decoding  performance using difference
> parameters by Kaldi.

That's great!

>        I have got some results, but I can't obtain an exact conclusion.
>        1) What is the biggest difference between  gmm-decode-faster and
> gmm-decode-kaldi? gmm-decode-kaldi can generate lattice, whilst
> gmm-decode-faster can't. I found the performances of gmm-decode-faster and
> gmm-decode-kaldi are almost same.

I think you mean the other way round, no?  Anyway the normal method to
do lattice
generation is gmm-latgen-faster, which is derived from
gmm-decode-faster.  gmm-decode-kaldi was
intended to be a highly optimized decoder, but in the end some of
those optimization techniques did
not work as well as we had hoped.

>        2) How do the decoding parameters of  gmm-decode-faster effect the
> performance? I found the acoustic-scale can obviously effect the
> performance, but other parameters almost have no effect.  In general,
> max-active=16000, beam=16.0, is the best one.

I would be cautious about saying that something is the best one.
Surely it depends
what point in the speed/accuracy tradeoff you want.

>        3) The beam in gmm-decode-faster is difference with that in Hvite in
> the scale.  How to adjust the beam to get an optimal result.

The difference is the acoustic scale.  The beam divided by the
acoustic scale should be
comparable to a HVite beam.

>        4) How to optimize the generated WFST to improve decoding
> performance? What are the  important matters needing to pay attention in
> composition, determination and minimization operations in Kaldi's WFST
> generation?

This is a very big subject.  There are many methods.
One is simply using a smaller WFST by using a smaller
language model, and then rescoring with a larger one.  There are two paths here:
either do lattice rescoring, or use the "biglm" decoder.  I would
appreciate it if you could
find out which approach is faster, as I have not had time.  In
general, I think such
experiments should take the form of generating speed/accuracy curves
and seeing which
curve lies above or below the other.

Search for "hbka.pdf" and read it, to learn the general principles and
theory of WFSTs-- although
that is a rather technical presentation.

Dan


> hlwang
>
>
>
> ------------------------------------------------------------------------------
> For Developers, A Lot Can Happen In A Second.
> Boundary is the first to Know...and Tell You.
> Monitor Your Applications in Ultra-Fine Resolution. Try it FREE!
> http://p.sf.net/sfu/Boundary-d2dvs2
> _______________________________________________
> Kaldi-users mailing list
> Kal...@li...
> https://lists.sourceforge.net/lists/listinfo/kaldi-users
>