Re: [Kaldi-developers] small grammar recognition from large grammar trained model

[Eamonn K. <Eam...@cs...>]

On 26/03/14 15:48, Daniel Povey wrote:
>
>     Is it possible to perform recognition confined to a small grammar
>     given
>     that you have trained on a large grammar that includes the small
>     grammar
>     as a subset?
>
>     I ask because I attempted to follow the recipe of
>     http://vpanayotov.blogspot.ie/2012/06/kaldi-decoding-graph-construction.html
>     to do this but to no avail.
>
>     Then I attempted to take egs/voxforge/s5/run.sh and strip out the
>     training section and change the corpus.txt file to obtain the small
>     grammar. The idea was that I would generate L and G using the existing
>     run script but then combine it with Ha and C to get the reduced
>     fst. It
>     all compiles and looks like it should work, but there must be a
>     mismatch
>     between the Ha of the existing large grammar model and the path
>     through
>     the model that uses the smaller G. The recogniser will respond to the
>     speaker but produces completely wrong results and in many cases just
>     produces the same word output every time.
>
>
> This should work.  I suspect you have done something like giving it 
> the wrong sample-rate audio.  The features are not comparable between 
> different sample rates. Check the log-likelihoods you get on decoding 
> (caution: they may or may not be printed out multiplied by the 
> acoustic scale)- if these are very different from your "matched" 
> decoding, then likely the acoustics are wrong.  Also see the fMLLR 
> objective function improvement, if you're using fMLLR- if the 
> acoustics are mismatched it will be very large, e.g. >5.
I used the online-gmm-decode-faster which is what I usually use. Giving 
the large grammar the same words it recognises them some of the time. 
Obviously my point of then using the small grammar is to improve the 
accuracy. I don't tend to use online-wav-gmm-decode-faster and I've 
found that if I use the sample rate of 44100 it gets flagged as an error 
anyway.

>
>     I've attempted also to see how I might use HCLG o G_s^- o G_l
>     where G_s
>     is the small grammar and G_l is the large grammar, but I see no
>     documentation on who this is actually performed using a script.
>
>
> This is implemented, it's called "biglm" in the code and scripts, 
> there is an example in the WSJ scripts, egs/wsj/s5/.
Thanks I'll have a look at it again. I was looking for it in the 
egs/wsj/s5 because I found a s3 version in bitbucket. Maybe I just 
missed it.
> Dan
>
>
>
>
>     Learn Graph Databases - Download FREE O'Reilly Book
>     "Graph Databases" is the definitive new guide to graph databases
>     and their
>     applications. Written by three acclaimed leaders in the field,
>     this first edition is now available. Download your free book today!
>     http://p.sf.net/sfu/13534_NeoTech
>     _______________________________________________
>     Kaldi-developers mailing list
>     Kal...@li...
>     <mailto:Kal...@li...>
>     https://lists.sourceforge.net/lists/listinfo/kaldi-developers
>
>


-- 
Best Regards,
    Eamonn

+                                   +                              +
   Eamonn Kenny B.A., M.Sc., Ph.D.     CNGL/Speech Communication Lab,
   Tel:   00+353-1-8961797             Dept. of Computer Science,
   Email: Eam...@sc...     F.34, O'Reilly Institute,
   http://www.cs.tcd.ie/Eamonn.Kenny   Trinity College Dublin,
   http://eamonnmkenny.wordpress.com   Dublin 2, Ireland.
+                                   +                              +